AI Role Analysis in Cybersecurity Sector

Cybersecurity as the name suggests is the process of safeguarding networks and programs from digital attacks. In today’s times, the world sustains on internet-connected systems that carry humungous data that is highly sensitive. Cyberthreats are on the rise with unscrupulous hackers taking over the entire industry by storm, with their unethical practices. This not only calls for more intense cyber security laws, but also the vigilance policies of the corporates, big and small, government as well as non-government; needs to be revisited.

With such huge responsibility being leveraged over the cyber-industry, more and more cyber-security enthusiasts are showing keen interest in the industry and its practices. In order to further the process of secured internet systems for all, unlike data sciences and other industries; the Cybersecurity industry has seen a workforce rattling its grey muscle with every surge they experience in cyber threats. Talking of AI impressions in Cybersecurity is still in its nascent stages of deployment as humans are capable of more; when assisted with the right set of tools.

Automatically detecting unknown workstations, servers, code repositories, and other hardware and software on the network are some of the tasks that could be easily managed by AI professionals, which were conducted manually by Cybersecurity folks. This leaves room for cybersecurity officials to focus on more urgent and critical tasks that need their urgent attention. Artificial intelligence can definitely do the leg work of processing and analyzing data in order to help inform human decision-making.

AI in cyber security is a powerful security tool for businesses. It is rapidly gaining its due share of trust among businesses for scaling cybersecurity. Statista, in a recent post, listed that in 2019, approximately 83% of organizations based in the US consider that without AI, their organization fails to deal with cyberattacks. AI-cyber security solutions can react faster to cyber security threats with more accuracy than any human. It can also free up cyber security professionals to focus on more critical tasks in the organization.


As it is said, “It takes a thief to catch a thief”. Being in its experimental stages, its cost could be an uninviting factor for many businesses. To counter the threats posed by cybercriminals, organizations ought to level up their internet security battle. Attacks backed by the organized crime syndicate with intentions to dismantle the online operations and damage the economy are the major threats this industry face today. AI is still mostly experimental and, in its infancy, hackers will find it much easy to carry out speedier, more advanced attacks. New-age automation-driven practices are sure to safeguard the crumbling internet security scenarios.


There are several advantageous reasons to embrace AI in cybersecurity. Some notable pros are listed below:

  •  Ability to process large volumes of data
    AI automates the creation of ML algorithms that can detect a wide range of cybersecurity threats emerging from spam emails, malicious websites, or shared files.
  • Greater adaptability
    Artificial intelligence is easily adaptable to contemporary IT trends with the ever-changing dynamics of the data available to businesses across sectors.
  • Early detection of novel cybersecurity risks
    AI-powered cybersecurity solutions can eliminate or mitigate the advanced hacking techniques to more extraordinary lengths.
  • Offers complete, real-time cybersecurity solutions
    Due to AI’s adaptive quality, artificial intelligence-driven cyber solutions can help businesses eliminate the added expenses of IT security professionals.
  • Wards off spam, phishing, and redundant computing procedures
    AI easily identifies suspicious and malicious emails to alert and protect your enterprise.


Alongside the advantages listed above, AI-powered cybersecurity solutions present a few drawbacks and challenges, such as:

  • AI benefits hackers
    Hackers can easily sneak into the data networks that are rendered vulnerable to exploitation.
  • Breach of privacy
    Stealing log-in details of the users and using them to commit cybercrimes, are deemed sensitive issues to the privacy of an entire organization.
  • Higher cost for talents
    The cost of creating an efficient talent pool is very high as AI-based technologies are in the nascent stage.
  • More data, more problems
    Entrusting our sensitive data to a third-party enterprise may lead to privacy violations.


AI professionals backed with the best AI certifications in the world assist corporations of all sizes to leverage the maximum benefits of the AI skills that they bring along, for the larger benefit of the organization. Cybersecurity teams and AI systems cannot work in isolation. This communion is a huge step forward to leveraging maximum benefits for secured cybersecurity applications for organizations. Hence, this makes AI in cybersecurity a much-coveted aspect to render its offerings in the long run.

Training of Deep Learning AI models

Alles dreht sich um Daten: die Trainingsmethoden des Deep Learning

Im Deep Learning gibt es unterschiedliche Trainingsmethoden. Welche wir in einem KI Projekt anwenden, hängt von den zur Verfügung gestellten Daten des Kunden ab: wieviele Daten gibt es, sind diese gelabelt oder ungelabelt? Oder gibt es sowohl gelabelte als auch ungelabelte Daten?

Nehmen wir einmal an, unser Kunde benötigt für sein Tourismusportal strukturierte, gelabelte Bilder. Die Aufgabe für unser KI Modell ist es also, zu erkennen, ob es sich um ein Bild des Schlafzimmers, Badezimmers, des Spa-Bereichs, des Restaurants etc. handelt. Sehen wir uns die möglichen Trainingsmethoden einmal an.

1. Supervised Learning

Hat unser Kunde viele Bilder und sind diese alle gelabelt, so ist das ein seltener Glücksfall. Wir können dann das Supervised Learning anwenden. Dabei lernt das KI Modell die verschiedenen Bildkategorien anhand der gelabelten Bilder. Es bekommt für das Training von uns also die Trainingsdaten mit den gewünschten Ergebnissen geliefert.
Während des Trainings sucht das Modell nach Mustern in den Bildern, die mit den gewünschten Ergebnissen zusammenpassen. So erlernt es Merkmale der Kategorien. Das Gelernte kann das Modell dann auf neue, ungesehene Daten übertragen und auf diese Weise eine Vorhersage für ungelabelte Bilder liefern, also etwa “Badezimmer 98%”.

2. Unsupervised learning

Wenn unser Kunde viele Bilder als Trainingsdaten liefern kann, diese jedoch alle nicht gelabelt sind, müssen wir auf Unsupervised Learning zurückgreifen. Das bedeutet, dass wir dem Modell nicht sagen können, was es lernen soll (die Zuordnung zu Kategorien), sondern es muss selbst Regelmäßigkeiten in den Daten finden.

Eine aktuell gängige Methode des Unsupervised Learning ist Contrastive Learning. Dabei generieren wir jeweils aus einem Bild mehrere Ausschnitte. Das Modell soll lernen, dass die Ausschnitte des selben Bildes ähnlicher zueinander sind als zu denen anderer Bilder. Oder kurz gesagt, das Modell lernt zwischen ähnlichen und unähnlichen Bildern zu unterscheiden.

Über diese Methode können wir zwar Vorhersagen erzielen, jedoch können diese niemals
die Ergebnisgüte von Supervised Learning erreichen.

3. Semi-supervised Learning

Kann uns unser Kunde eine kleine Menge an gelabelten Daten und eine große Menge an nicht gelabelten Daten zur Verfügung stellen, wenden wir Semi-supervised Learning an. Diese Datenlage begegnet uns in der Praxis tatsächlich am häufigsten. Bei fast allen KI Projekten stehen einer kleinen Menge an gelabelten Daten ein Großteil an unstrukturierten
Daten gegenüber.

Mit Semi-supervised Learning können wir beide Datensätze für das Training verwenden. Das gelingt zum Beispiel durch die Kombination von Contrastive Learning und Supervised Learning. Dabei trainieren wir ein KI Modell mit den gelabelten Daten, um Vorhersagen für Raumkategorien zu erhalten. Gleichzeitig lassen wir es Ähnlichkeiten und Unähnlichkeiten in den ungelabelten Daten erlernen und sich daraufhin selbst optimieren. Auf diese Weise können wir letztendlich auch gute Label-Vorhersagen für neue, ungesehene Bilder erzielen.

Fazit: Supervised vs. Unsupervised vs. Semi-supervised

Supervised Learning wünscht sich jeder, der mit einem KI Projekt betraut ist. In der Praxis ist das kaum anwendbar, da selten sämtliche Trainingsdaten gut strukturiert und gelabelt vorliegen.

Wenn nur unstrukturierte und ungelabelte Daten vorhanden sind, dann können wir mit Unsupervised Learning immerhin Informationen aus den Daten gewinnen, die unser Kunde so nicht hätte. Im Vergleich zu Supervised Learning ist aber die Ergebnisqualität deutlich schlechter.

Mit Semi-Supervised Learning versuchen wir das Datendilemma, also kleiner Teil gelabelte, großer Teil ungelabelte Daten, aufzulösen. Wir verwenden beide Datensätze und können gute Vorhersage-Ergebnisse erzielen, deren Qualität dem Supervised Learning oft ebenbürtig sind.

Dieser Artikel entstand in Zusammenarbeit zwischen DATANOMIQ, einem Unternehmen für Beratung und Services rund um Business Intelligence, Process Mining und Data Science. und pixolution, einem Unternehmen für AI Solutions im Bereich Computer Vision (Visuelle Bildsuche und individuelle KI Lösungen).

Stop saying “trial and errors” for now: seeing reinforcement learning through some spectrums

*This is the fourth article of the series My elaborate study notes on reinforcement learning.

*In this article series “the book by Barto and Sutton” means “Reinforcement Learning: An Introduction second edition.” This book is said to be almost mandatory for those who seriously learn Reinforcement Learning (RL). And “the whale book” means a Japanese textbook named 「強化学習 (機械学習プロフェッショナルシリーズ)」(“Reinforcement Learning (Machine Learning Processional Series)”), by Morimura Tetsuro. I would say the former is for those who want to mainly learn how to use RL, and the latter is for more theoretical understanding. I am trying to make something between them in my series.

1, Finally to reinforcement learning

Some of you might have got away with explaining reinforcement learning (RL) only by saying an obscure thing like “RL enables computers to learn through trial and errors.” But if you have patiently read my articles so far, you might have come to say “RL is a family of algorithms which simulate procedures similar to dynamic programming (DP).” Even though my article series has not covered anything concrete and unique to RL yet, I think my series has already laid a hopefully effective foundation of discussions on RL. And in the first article, I already explained that “trial and errors” are only agents’ actions for collecting data from the environment. Such “trial and errors” lead to “experiences” of computers. And in this article we can finally start discussing how computers “experience” things in more practical and theoretical ways.

*The expression “to learn” is also frequently used in contexts of other machine learning algorithms. Thus in order to clearly separate the ideas, let me use the expression “to experience” when it comes to explaining RL. At any rate, what computers are doing is updating parameters, and in RL also updating values and policies. But some terms related to RL also use the word “experience,” for example experience replay, so “to experience” might be a preferred phrase in RL fields.

I think changing discussions on DP into those on RL is like making graphs more “open” rather than “closed.” In the second article, I explained DP problems, where the models of environments are completely known, as repeatedly updating graphs like neural networks. As I have been repeatedly saying RL, or at least model-free RL, is an approximated application of DP in the environments without a complete model. That means, connections of nodes of the graph, that is relations of actions and states, are something agents have to estimate directly or indirectly. I think that can be seen as untying connections of the graphs which I displayed when I explained DP. By doing so, I propose to see RL or more exactly model-free RL like the graph of the right side of the figure below.

*For the time being, I would prefer to use the term model-free RL rather than just RL. That is not only because this article is about model-free RL but also because I want to avoid saying inaccurate things about wider range of RL algorithms I would have to study more precisely and explain.

Some people might say these are tree structures, and that might be technically correct. But in my sense, this is more of “willows.” The cover of the second edition of the books by Barto and Sutton also looks like willows. The cover design comes from a paper on RL named “Learning to Drive a Bicycle using Reinforcement Learning and Shaping.” The paper is about learning to ride a bike in a simulator with RL. The geometric patterns are not models of human brain nerves, but trajectories of an agent learning to balance a bike. However interestingly, the trajectories of the bike, which are inscribed on a road, partly diverge but converge in a certain way as a whole, like the RL graph I propose. That is why I chose some pictures of 「花札 (hanafuda)」as the main picture of this series. Hanafuda is a Japanese gamble card game with monthly seasonal flower pictures. And the cards of June have pictures of willows.

Source: Learning to Drive a Bicycle using Reinforcement Learning and Shaping, Randløv, (1998)    Richard S. Sutton, Andrew G. Barto, “Reinforcement Learning: An Introduction,” MIT Press, (2018)

2, Untying DP graphs: planning or learning

Even though I have just loudly declared that my RL graphs are more of “willow” structures in my aesthetic sense, I must admit they should basically be discussed as popular tree structures. That is because, when you start discussing practical RL algorithms you need to see relations of states and actions as tree structures extending. If you already more or less familiar with tree structures or searching algorithms on tree graphs, learning RL with tree structures should be more or less straightforward to you. Another reason for using tree structures with nodes of states and actions is that the book by Barto and Sutton use buck up diagrams of Bellman equations which are tree graphs. But I personally think the graphs should be used more effectively, so I am trying to expand its uses to DP and RL algorithms in general. In order to avoid confusions about current discussions on RL in my article series, I would like to give an overall review on how to look at my graphs.

The graphs in the figure below are going to be used in my articles, at least when I talk about model-free RL. I made them based on the backup diagram of Bellman equation introduced in the book by Barto and Sutton. I would like you to first remember that in RL we are basically discussing Markov decision process (MDP) environment, where the next action and the resulting next states depends only on the current state. Such models are composed of white nodes representing each state s in an state space \mathcal{S}, and black nodes representing each action a, which is a member of an action space \mathcal{A}. Any behaviors of agents are represented as going back and forth between black and white nodes of the model, and that is why connections in the MDP model are bidirectional.  In my articles let me call such model of environments “a closed model.” RL or general planning problems are matters of optimizing policies in such models of environments. Optimizing the policies are roughly classified into two types, planning/searching or RL, and the main difference between them is whether connections of graphs of models are known or not. Planning or searching is conducted without actually moving in the environment. DP are family of planning algorithms which are known to converge, and so far in my articles we have seen that DP are enabled by repeatedly applying Bellman operators. But instead of considering and updating all the possible transitions in the model like DP, planning can be conducted more sparsely. Such sparse planning are often called searching, and many of them use tree structures. If you have learned any general decision making problems with tree graphs, you might be already familiar with some searching techniques like alpha-beta pruning.

*In explanations on DP in my articles, directions of connections of model graphs are confusing, so I precisely explained how to look at them in the second section in the last article.

On the other hand, RL algorithms are matters of learning the linkages of models of environments by actually moving in them. For example, when the agent in the figure below move on a grid map like the purple arrows, the movement is represented like in the closed model in the middle. However as the agent does not have the complete closed model, the agent has to move around in the environment like the tree structure at the right side to learn values of each node.

The point is, whether models of environments are known or unknown, or whether agents actually move in the environment or not, movements of agents are basically represented as going back and forth between white nodes and black nodes in closed models. And such closed models are entangled in searching or RL. They are similar operations, but they are essentially different in that searching agents do not actually move in searching but in RL they actually move.  In order to distinguish searching and learning, in my articles, trees for searching are extended vertically, trees for learning horizontally.

*DP and searching are both planning, but DP consider all the connections of actions and states by repeatedly applying Bellman operators. Thus I would not count DP as “untying” of closed models.

3, Some spectrums in RL algorithms

Starting studying actual RL algorithms also means encountering various algorithms one after another. Some of you might have already been overwhelmed by new terms coming up one after another in study materials on RL. That is because, as I explained in the first article, RL is more about how to train models of values or policies. Thus it is natural that compared to general machine learning, which more or less share the same training frameworks, RL has a variety of training procedures. Rather than independently studying each RL algorithm, I think it is more effective to see connections of each algorithm, which is linked by adjusting degrees of some important elements in RL. In fact I have already introduced those elements as some pairs of key words of RL in the first article. But it would be all the more effective to review them, especially after learning DP algorithms as representative planning methods. If you study RL that way, you would come to see trial and errors or RL as a crucial but just one aspect of RL.

I think if you care less about the trial-and-error aspect of RL that allows you to study RL more effectively in the beginning. And for the time being, you should stop viewing RL in the popular way as presented above. Not that I am encouraging you to ignore the trial and error part, namely relations of actions, rewards, and states. My point is that it is more of inside the agent that should be emphasized. Planning, including DP is conducted inside the agent, and trial and errors are collection of data from the environment for the sake of the planning. That is why in many study materials on RL, DP is first introduced. And if you see differences of RL algorithms as adjusting of some pairs of elements of planning problems, it would be less likely that you would get lost in curriculums on RL. The pairs are like some spectrums. Not that you always have to choose either of each pair, but rather ideal solutions are often in the middle of the two ends of the spectrums depending on tasks. Let’s take a look at the types of those spectrums one by one.

(1) Value-policy or actor-critic spectrum

The crucial type of spectrum you should be already familiar with is the value-policy one. I think this spectrum can be adjusted in various ways. For example, over the last two articles we have seen how values and policies reach the optimal functions in DP using policy iteration or value iteration. Policy iteration alternates between updating values and policies until convergence to the optimal policy, whereas value iteration keeps updating only values until reaching the optimal value, to get the optimal policy at the end. And similar discussions can be seen also in the upcoming RL algorithms. The book by Barto and Sutton sees such operations in general as generalized policy iteration (GPI).

Source: Richard S. Sutton, Andrew G. Barto, “Reinforcement Learning: An Introduction,” MIT Press, (2018)

You should pay attention to the idea of GPI because this is what makes RL different form other general machine learning. In many cases RL is explained as a field of machine learning which is like trial and errors, but I personally think that GPI, interactive optimization between values and policies, should be more emphasized. As I said in the first article, RL optimizes decision making rules, that is policies \pi(a|s), in MDPs. Other general machine learning algorithms have more direct supervision by loss functions and models are optimized so that loss functions are minimized. In the case of the figure below, an ML model f is optimized to f_{\ast} by optimization such as gradient descent. But on the other hand in RL policies \pi do not have direct loss functions. Then RL uses values v(s), which are functions of how good it is to be in states s. As one part of GPI, the value function v_{\pi} for the current policy \pi is calculated, and this is called estimation in the book by Barto and Sutton.  And based on the estimated value function, the policy is improved as \pi ', which is called policy improvement, and overall processes of estimation and policy improvement are called control in the book. And v_{\pi} and \pi are updated alternately this way until converging to the optimal values v_{\ast} or policies \pi_{\ast}. This interactive updates of values and policies are done inside the agent, in the dotted frame in red below. I personally think this part should be more emphasized than trial-and-error-like behaviors of agents. Once you see trial and errors of RL as crucial but just one aspect of GPI and focus more inside agents, you would see why so many study materials start explaining RL with DP.

You can explicitly model such interactions of values and policies by modeling each of them with different functions, and in this case such frameworks of RL in general are called actor-critic methods. I am gong to explain actor-critic methods in an upcoming article. Thus the value-policy spectrum also can be seen as a actor-critic spectrum. Differences between the pairs of value-policy or actor-critic spectrums are something you would little by little understand. For now I would say GPI is the most general and important idea behind RL. But practical RL algorithms are implemented as actor-critic methods. Critic parts gives some signals to actor parts, and critic parts get its consequence by actor parts taking actions in environments. Not that actors directly give feedback to critics.

*I think one of confusions in studying RL come from introducing Q-learning or SARSA at the first algorithms or a control in RL. As I have said earlier, interactive relations between values and policies or actors and critics, that is GPI, should be emphasized. And I think that is why DP is first introduced in many books. But in Q-learning or SARSA, an actor and a critic parts are combined as one module. But explicitly separating the actor and critic parts would be just too difficult at the beginning. And modeling an actor and a critic with separate modules would lead to difficulties in optimizing them together.

(2) Exploration-exploitation or on-off policy spectrum

I think the most straightforward spectrum is the exploitation-exploration spectrum. You can adjust how likely agents take random actions to collect data. Occasionally it is ideal for agents to have some degree of randomness in taking actions to explore unknown states of environments. One of the simplest algorithms to formulate randomness of actions is ε-greedy method, which I explained in the first article. In this method in short agents take a random action with a probability of ε. Instead of arbitrarily setting a hyperparameter \epsilon, randomness of actions can be also learned by modeling policies with certain functions. This randomness of functions can be also modeled in actor-critic frameworks. That means, depending on a choice of an actor, such actor can learn randomness of actions, that is explorations.

The two types of spectrums I have introduced so far lead to another type of spectrum. It is an on-off policy spectrum. Even though I explained types of policies in the last article using examples of home-lab-Starbucks diagrams, there is another way to classify policies: there are target policies and behavior policies. The former are the very policies whose optimization we have been discussing. The latter are policies for taking actions and collecting data. When agents use target policies also as behavior policies, they are on-policy algorithms. If agents use different policies for taking actions during optimization of target policies, they are off-policy methods.

Policy iteration and value iteration of DP can be also classified into on-policy or off-policy in a sense. In policy iteration values are updated using an up-to-date estimated policy, and the policy becomes optimal when it converges. Thus behavior and target policies are the same in this case. On the other hand in value iteration, values are updated with Bellman optimality operator, which updates values in a greedy way. Using greedy method means the policy \pi is not used for considering which action to take. Thus target and behavior policies are different. As you will see soon, concrete model-free RL algorithms like SARSA or Q-learning also have the same structure: the former is on-policy and the latter is off-policy. The difference of on-policy or off-policy would be more straightforward if we model behavior policies and target policies with different functions. An advantage of off-policy RL is you can model randomness of exploration of agents with extra functions. On the other hand, a disadvantage is that it would be harder to train different models at the same time. That might be a kind of tradeoff similar to an actor-critic method.

Even though this exploration-exploitation aspect of RL is relatively easy to understand, at the same time that can lead to much more complicated discussions on RL, which I would not be able to cover in this article series. I recommended you to stop seeing RL as trial and errors for the time being, but in the end trial and errors would prove to be crucial because data needed for GPI are collected mainly via trial and errors. Even if you implement some simple RL algorithms, you would soon realize it is hard to deal with unvisited states. Enough explorations need to be modeled by a behavior policy or some sophisticated heuristic techniques. I am planning to explain convergence of several RL algorithms, and they are guaranteed by sufficiently exploring all the states. However, thorough explorations of all the states lead to massive computational costs. But lack of exploration would let RL agents myopically overestimate current policies, never finding policies which pay off in the long run. That might be close to discussions on how to efficiently find a global minimum of a loss function, avoiding local minimums.

(3) TD-MonteCarlo spectrum

A variety of spectrums so far are enabled by modeling proper functions on demand. But in AI problems such functions are something which have to be automatically trained with some supervision. Instead of giving supervision explicitly with annotated data like in supervised learning of general machine learning, RL agents train models with “experiences.” As I am going to explain in the next part of this article, “experiences” in RL contexts mean making some estimations of values and adjusting such estimations based on actual rewards they get. And the timings of such feedback lead to another spectrum, which I call a TD-MonteCarlo spectrum. When the feedback happens every time an agent takes an action, it is TD method, on the other hand when that happens only at the end of an episode, that is Monte Carlo method. But it is easy to imagine that ideal solutions are usually at the middle of them. I am going to dig this topic soon in the next article. And n-step methods or TD(λ), which bridge the TD and Monte Carlo, are going to be covered in one of upcoming articles.

(4) Model free-based spectrum

The next spectrum might be relatively hard to understand, and to be honest I am still not completely sure about this topic. Please bear that in your mind. In the last section, I said RL is a kind of untying DP graphs and make them open because in RL, models of environments are unknown. However to be exact, that was mainly about model-free RL, which this article is going to cover for the time being. And I would say the graphs I showed in the last section were just two extremes of this model based-free spectrum. Some model-based RL methods exist in the middle of those two ends. In short RL agents can retain models of environments and do some plannings even when they do trial and errors. The figure below briefly compares planning, model-based RL, and model-free RL in the spectrum.

Let’s take a rough example of humans solving a huge maze. DP, which I have covered is like having a perfect map of the maze and making plans of how to move inside in advance. On the other hand, model-free reinforcement learning is like soon actually entering the maze without any plans. In model-free reinforcement learning, you only know how big the maze is, and you have a great memory for remembering in which directions to move, in all the places. However, as the model of how paths are connected is unknown, and you naively try to remember all the actions in all the places, it generally takes a longer time to solve the maze. As you could easily imagine, having some heuristic ideas about the model of the maze and taking some notes and making plans about courses would be the most efficient and the most peaceful. And such models in your head can be updated by actually moving in the maze.

*I believe that you would not say the pictures above are spoilers.

I need to more clearly talk about what a model is in RL or general planning problems. The book by Barto and Sutton simply defines a model this way: “By a model of the environment we mean anything that an agent can use to predict how the environment will respond to its actions. ” The book also says such models can be also classified to distribution models and sample models. The difference between them is the former describes an environment as combinations of known models, but the latter is like a black box model of an environment. An intuitive example is, as introduced in the book by Barto and Sutton, throwing dozens of dices can be seen in the both types. If you just throw the dices, sometimes chancing numbers of dices, and record the sum of the numbers on the dices s every time, that is equal to getting the sum from a black box. But a probabilistic distribution of such sums can be actually calculated as a multinomial distribution. Just as well, you can see a probability of transitions in an RL environment as a black box, but the probability can be also modeled. Some readers might have realized that distribution or sample models can be almost the same in the end, with sufficient data. In many cases of machine learning or statistics algorithms, complicated distributions have to be approximated with samples. Or rather how to approximate them is more of interest. In the case of dozens of dices, you can analytically calculate its distribution model as a multinomial distribution. But if you throw the dices numerous times, you would get precise approximated distributions.

When we discuss model-based RL, we need to consider not only DP but also other planning algorithms. DP is a family of planning algorithms which are known to converge, and many of RL algorithms share a lot with DP at theoretical levels. But in fact DP has one shortcoming even if the MDP model of an environment is known: DP needs to consider and update all the states. When models of environments are too complicated and large, applying DP is not a good idea. Also in many of such cases, you could not even get such a huge model of the environment. You would rather get only a black box model of the environment. Such a black box model only gets a pair of current state and action (s, a), and gives out the next state s' and corresponding reward r, that is the black box is a sample model. In this case other planning methods with some searching algorithms are used, for example Monte Carlo tree search. Such search algorithms are designed to more efficiently and sparsely search states and actions of interest. Many of searching algorithms used in RL make uses of tree structures. Model-based approaches can be roughly classified into three types below based on size or complication of models.

*As you could see, differences between sample models and distribution models can be very ambiguous. So are differences between model-free and model-based RL, I guess. As a matter of fact the whale book says the distributions of models approximated in model-free RL are the same as those in model-based ones. I cannot say anything exactly anymore, but I guess model-free RL is more of “memorizing” an environment, or combinations of states and actions in the environments. But memorizing environments can be computationally problematic in many cases, so assuming some distributions of models can help. That is my impression for now.

*Tree search algorithms alone shows very impressive performances, as long as you have massive computation resources. A heuristic tree search without reinforcement learning could defeat Garri Kasparow, a former chess champion, as long as enough computation resource is available. Searching algorithms were enough for “simplicity” of chess.

*I am not sure whether model-free RL algorithms are always simpler than model-based ones. For example Deep Q-Learning, a model-free method with some neural networks can learn to play Atari or Nintendo Entertainment System. Model-based deep RL is used in more complex task like AlphaGo or AlphaZero, which can defeat world champions of various board games. AlphaGo or AlphaZero models intuitions in phases of board games with convolutional neural networks (CNN), prediction of some phases ahead with search algorithms, and learning from past experiences with RL. I am not going to cover model-based RL in general in this series, but instead I would like to explain how RL enables computers to play video games after introducing some searching algorithms.

(5) Model expressivity spectrum

No matter how impressive or dreamy RL algorithms sound, their competence largely depend on model expressivity. In the first article, I emphasized “simplicity” of RL. DP or RL algorithms so far or in upcoming several articles consider incredibly simple cases like kids playbooks. And that beginning parts of most RL study materials cover only the left side of the figure below. In order to enable RL agents with more impressive tasks such as balancing cart-pole or playing video games, we need to raise the bar of expressivity spectrum, from the left to the right side of the figure below. You need to wait until a chapter or a section on “function approximation” in order to actually feel that your computer is doing trial and errors. And such chapters finally appear after reading half of both the book by Barto and Sutton and the whale book.

*And this spectrum is also a spectrum of computation costs or convergence. The left type could be easily implemented like programming assignments of schools since it in short needs only Excel sheets, and you would soon get results. The middle type would be more challenging, but that would not b computationally too expensive. But when it comes to the type at the right side, that is not something which should be done on your local computer. At least you need a GPU. You should expect some hours or days even for training RL agents to play 8 bit video games. That is of course due to cost of training deep neural networks (DNN), especially CNN. But another factors is potential inefficiency of RL. I hope I could explain those weak points of RL and remedies for them.

We need to model values and policies with certain functions. For the time being, in my articles values and policies are just modeled as tabular data, that is some NumPy arrays or Excel sheets. These are types of cases where environments and actions are relatively simple and discrete. Thus they can be modeled with some tabular data with the same degree of freedom. Assume a case where there are only 30 grids in an environment and only 4 types of actions in every grid. In such case, values are stored as arrays with 30 elements, and so are policies. But when environments are more complex or require continuous values of some parameters, values and policies have to be approximated with some models. When only relatively few parameters need to be estimated, simple machine learning models such as softmax functions can be used as such models. But compared to the cases with tabular data, convergence of training has to be discussed more carefully. And when you need to estimate continuous values, techniques like policy gradients have to be introduced. And we can dramatically enhance expressivity of models with deep neural netowrks (DNN), and such RL is called deep RL. Deep RL has showed great progress these days, and it is capable of impressive performances. Deep RL often needs observers to process inputs like video frames, and for example convolutional neural networks (CNN) can be used to make such observers. At any rate, no matter how much expressivity RL models have, they need to be supervised with some signals just as general machine learning often need labeled data. And “experiences” give such supervisions to RL agents.

(6) Adjusting sliders of spectrum

As you might have already noticed, these spectrums are not something you can adjust independently like faders on mixing board. They are more like some sliders for adjusting colors, brightness, or chroma on painting software. If you adjust one element, other parts are more or less influenced. And even though there are a variety of colors in the world, they continuously change by adjusting those elements of colors. Just as well, even if each RL algorithms look independent, many of them share more or less the same ideas, and only some parts are different in terms of their degrees. When you get lost in the course of studying RL, I would like you to decompose the current topic into these spectrums of RL elements I have explained.

I hope my explanations so far changed how you see RL. In the first article I already said RL is approximation of DP-like procedures with data collected by trial and errors, but from now on I would explain it also this way: RL is a family of algorithms which enable GPI by adjusting some spectrums.

In the next some articles, I am going to mainly cover RL algorithms named SARSA and Q-learning. Both of them use tabular data, and they are model-free. And in values and policies, or actors and critics are together modeled as action-value functions, which I am going to explain later in this article. The only difference is SARSA is on-policy, and Q-learning is off-policy, just as I have already mentioned. And when it comes to how to train them, they both use Temporal Difference (TD), and this gives signals of “experience” to RL agents. Altering DP in to model-free RL is, in the figure above, adjusting the model-based-free and MonteCarlo-TD spectrums to the right end. And you also adjust the low-high-expressivity and value-policy spectrums to the left end. In terms of actor-critic spectrum, the actor and the critic parts are modeled as the same module. Seeing those algorithms this way would be much more effective than looking at their pseudocode independently.

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Training of Deep Learning AI models

It’s All About Data: The Training of AI Models

In deep learning, there are different training methods. Which one we use in an AI project depends on the data provided by our customer: how much data is there, is it labeled or unlabeled? Or is there both labeled and unlabeled data?

Let’s say our customer needs structured, labeled images for an online tourism portal. The task for our AI model is therefore to recognize whether a picture is a bedroom, bathroom, spa area, restaurant, etc. Let’s take a look at the possible training methods.

1. Supervised Learning

If our customer has a lot of images and they are all labeled, this is a rare stroke of luck. We can then apply supervised learning. The AI model learns the different image categories based on the labeled images. For this purpose, it receives the training data with the desired results from us.

During training, the model searches for patterns in the images that match the desired results, learning the characteristics of the categories. The model can then apply what it has learned to new, unseen data and in this way provide a prediction for unlabeled images, i.e., something like “bathroom 98%.”

2. Unsupervised Learning

If our customer can provide many images as training data, but all of them are not labeled, we have to resort to unsupervised learning. This means that we cannot tell the model what it should learn (the assignment to categories), but it must find regularities in the data itself.

Contrastive learning is currently a common method of unsupervised learning. Here, we generate several sections from one image at a time. The model should learn that the sections of the same image are more similar to each other than to those of other images. Or in short, the model learns to distinguish between similar and dissimilar images.

Although we can use this method to make predictions, they can never achieve the quality of results of supervised learning.

3. Semi-supervised Learning

If our customer can provide us with few labeled data and a large amount of unlabeled data, we apply semi-supervised learning. In practice, we actually encounter this data situation most often.

With semi-supervised learning, we can use both data sets for training, the labeled and the unlabeled data. This is possible by combining contrastive learning and supervised learning, for example: we train an AI model with the labeled data to obtain predictions for room categories. At the same time, we let the model learn similarities and dissimilarities in the unlabeled data and then optimize itself. In this way, we can ultimately achieve good label predictions for new, unseen images.

Supervised vs. Unsupervised vs. Semi-supervised

Everyone who is entrusted with an AI project wants to apply supervised learning. In practice, however, this is rarely the case, as rarely all training data is well structured and labeled.

If only unstructured and unlabeled data is available, we can at least extract information from the data with unsupervised learning. These can already provide added value for our customer. However, compared to supervised learning, the quality of the results is significantly worse.

With semi-supervised learning, we try to resolve the data dilemma of small part labeled data, large part unlabeled data. We use both datasets and can obtain good prediction results whose quality is often on par with those of supervised learning. This article is written in cooperation between DATANOMIQ and pixolution, a company for computer vision and AI-bases visual search.

Automatic Financial Trading Agent for Low-risk Portfolio Management using Deep Reinforcement Learning

This article focuses on autonomous trading agent to solve the capital market portfolio management problem. Researchers aim to achieve higher portfolio return while preferring lower-risk actions. It uses deep reinforcement learning Deep Q-Network (DQN) to train the agent. The main contribution of their work is the proposed target policy.


Author emphasizes the importance of low-risk actions for two reasons: 1) the weak positive correlation between risk and profit suggests high returns can be obtained with low-risk actions, and 2) customer satisfaction decreases with increases in investment risk, which is undesirable. Author challenges the limitation of Supervised Learning algorithm since it requires domain knowledge. Thus, they propose Reinforcement Learning to be more suitable, because it only requires state, action and reward specifications.

The study verifies the method through the back-test in the cryptocurrency market because it is extremely volatile and offers enormous and diverse data. Agents then learn with shorter periods and are tested for the same period to verify the robustness of the method. 

2 Proposed Method

The overall structure of the proposed method is shown below.

The architecutre of the proposed trading agent system.

The architecutre of the proposed trading agent system.

2.1 Problem Definition

The portfolio consists of m assets and one base currency.

The price vector p stores the price p of all assets:

The portfolio vector w stores the amount of each asset:

At time 𝑡, the total value W_t of the portfolio is defined as the inner product of the price vector p_t and the portfolio vector w_t .

Finally, the goal is to maximize the profit P_t at the terminal time step 𝑇.

2.2 Asset Data Preprocessing

1) Asset Selection
Data is drawn from the Binance Exchange API, where top m traded coins are selected as assets.

2) Data Collection
Each coin has 9 properties, shown in Table.1, so each trade history matrix has size (α * 9), where α is the size of the target period converted into minutes.

3) Zero-Padding
Pad all other coins to match the matrix size of the longest coin. (Coins have different listing days)

Comment: Author pointed out that zero-padding may be lacking, but empirical results still confirm their method covering the missing data well.

4) Stack Matrices
Stack m matrices of size (α * 9) to form a block of size (m* α * 9). Then, use sliding window method with widow size w to create (α – w + 1) number of sequential blocks with size (w *  m * 9).

5) Normalization
Normalize blocks with min-max normalization method. They are called history block 𝜙 and used as input (ie. state) for the agent.

3. Deep Q-Network

The proposed RL-based trading system follows the DQN structure.

Deep Q-Network has 2 networks, Q- and Target network, and a component called experience replay. The Q-network is the agent that is trained to produce the optimal state-action value (aka. q-value).

Comment: Q-value is calculated by the Bellman equation, which, in short, consists of the immediate reward from next action, and the discounted value of the next state by following the policy for all subsequent steps.


Agent: Portfolio manager
Action a: Trading strategy according to the current state
State 𝜙 : State of the capital market environment
Environment: Has all trade histories for assets, return reward r and provide next state 𝜙’ to agent again

DQN workflow:

DQN gets trained in multiple time steps of multiple episodes. Let’s look at the workflow of one episode.

Training of a Deep Q-Network

Training of a Deep Q-Network

1) Experience replay selects an action according to the behavior policy, executes in the environment, returns the reward and next state. This experience set (\phi_t, a_t, r_r,\phi_{t+!}) is stored in the repository as a sample of training data.

2) From the repository of prior observations, take a random batch of samples as the input to both Q- and Target network. The Q-network takes the current state and action from each data sample and predicts the q-value for that particular action. This is the ‘Predicted Q-Value’.Comment: Author uses 𝜀-greedy algorithm to calculate q-value and select action. To simplify, 𝜀-greedy policy takes the optimal action if a randomly generated number is greater than 𝜀, which represents a tradeoff between exploration and exploitation.

The Target network takes the next state from each data sample and predicts the best q-value out of all actions that can be taken from that state. This is the ‘Target Q-Value’.

Comment: Author proposes a different target policy to calculate the target q-value.

3) The Predicted q-value, Target q-value, and the observed reward from the data sample is used to compute the Loss to train the Q-network.

Comment: Target Network is not trained. It is held constant to serve as a stable target for learning and will be updated with a frequency different from the Q-network.

4) Copy Q-network weights to Target network after n time steps and continue to next time step until this episode is finished.

The architecutre of the proposed trading agent system.

4.0 Main Contribution of the Research

4.1 Action and Reward

Agent determines not only action a but ratio , at which the action is applied.

  1. Action:
    Hold, buy and sell. Buy and sell are defined discretely for each asset. Hold holds all assets. Therefore, there are (2m + 1) actions in the action set A.

    Agent obtains q-value of each action through q-network and selects action by using 𝜀-greedy algorithm as behavior policy.
  2. Ratio:
    \sigma is defined as the softmax value for the q-value of each action (ie. i-th asset at \sigma = 0.5 , then i-th asset is bought using 50% of base currency).
  3. Reward:
    Reward depends on the portfolio value before and after the trading strategy. It is clipped to [-1,1] to avoid overfitting.

4.2 Proposed Target Policy

Author sets the target based on the expected SARSA algorithm with some modification.

Comment: Author claims that greedy policy ignores the risks that may arise from exploring other outcomes other than the optimal one, which is fatal for domains where safe actions are preferred (ie. capital market).

The proposed policy uses softmax algorithm adjusted with greediness according to the temperature term 𝜏. However, softmax value is very sensitive to the differences in optimal q-value of states. To stabilize  learning, and thus to get similar greediness in all states, author redefine 𝜏 as the mean of absolute values for all q-values in each state multiplied by a hyperparameter 𝜏’.

4.3 Q-Network Structure

This study uses Convolutional Neural Network (CNN) to construct the networks. Detailed structure of the networks is shown in Table 2.

Comment: CNN is a deep neural network method that hierarchically extracts local features through a weighted filter. More details see:

5 Experiment and Hyperparameter Tuning

5.1 Experiment Setting

Data is collected from August 2017 to March 2018 when the price fluctuates extensively.

Three evaluation metrics are used to compare the performance of the trading agent.

  • Profit P_t introduced in 2.1.
  • Sharpe Ratio: A measure of return, taking risk into account.

    Comment: p_t is the standard deviation of the expected return and P_f  is the return of a risk-free asset, which is set to 0 here.
  • Maximum Drawdown: Maximum loss from a peak to a through, taking downside risk into account.

5.2 Hyperparameter Optimization

The proposed method has a number of hyperparameters: window size mentioned in 2.2,  𝜏’ in the target policy, and hyperparameters used in DQN structure. Author believes the former two are key determinants for the study and performs GridSearch to set w = 30, 𝜏’ = 0.25. The other hyperparameters are determined using heuristic search. Specifications of all hyperparameters are summarized in the last page.

Comment: Heuristic is a type of search that looks for a good solution, not necessarily a perfect one, out of the available options.

5.3 Performance Evaluation

Benchmark algorithms:

UBAH (Uniform buy and hold): Invest in all assets and hold until the end.
UCRP (Uniform Constant Rebalanced Portfolio): Rebalance portfolio uniformly for every trading period.

Methods from other studies: hyperparameters as suggested in the studies
EG (Exponential Gradient)
PAMR (Passive Aggressive Mean Reversion Strategy)

Comment: DQN basic uses greedy policy as the target policy.

The proposed DQN method exhibits the best overall results out of the 6 methods. When the agent is trained with shorter periods, although MDD increases significantly, it still performs better than benchmarks and proves its robustness.

6 Conclusion

The proposed method performs well compared to other methods, but there is a main drawback. The encoding method lacked a theoretical basis to successfully encode the information in the capital market, and this opaqueness is a rooted problem for deep learning. Second, the study focuses on its target policy, while there remains room for improvement with its neural network structure.

Specification of Hyperparameters

Specification of Hyperparameters.



  1. Shin, S. Bu and S. Cho, “Automatic Financial Trading Agent for Low-risk Portfolio Management using Deep Reinforcement Learning”,
  2. Li, P. Zhao, S. C. Hoi, and V. Gopalkrishnan, “PAMR: passive aggressive mean reversion strategy for portfolio selection,” Machine learning, vol. 87, pp. 221-258, 2012.
  3. P. Helmbold, R. E. Schapire, Y. Singer, and M. K. Warmuth, “On‐line portfolio selection using multiplicative updates,” Mathematical Finance, vol. 8, pp. 325-347, 1998.,can%20be%20interpreted%20as%20probabilities.

Training of Deep Learning AI models

Ein KI Projekt richtig umsetzen : So geht’s

Sie wollen in Ihrem Unternehmen Kosten senken und effizientere Workflows einführen? Dann haben Sie vielleicht schon darüber nachgedacht, Prozesse mit Künstlicher Intelligenz zu automatisieren. Für einen gelungenen Start, besprechen wir nun, wie ein KI-Projekt abläuft und wie man es richtig umsetzt.

Wir von DATANOMIQ und pixolution teilen unsere Erfahrungen aus Deep Learning Projekten, wo es vor allem um die Optimierung und Automatisierung von Unternehmensprozessen rund um visuelle Daten geht, etwa Bilder oder Videos. Wir stellen Ihnen die einzelnen Projektschritte vor, verraten Ihnen, wo dabei die Knackpunkte liegen und wie alle Beteiligten dazu beitragen können, ein KI-Projekt zum Erfolg zu führen.

1. Erstgespräch

In einem Erstgespräch nehmen wir Ihre Anforderungen auf.

  • Bestandsaufnahme Ihrer aktuellen Prozesse und Ihrer Änderungswünsche: Wie sind Ihre aktuellen Prozesse strukturiert? An welchen Prozessen möchten Sie etwas ändern?
  • Zielformulierung: Welches Endergebnis wünschen Sie sich? Wie genau sollen die neuen Prozesse aussehen? Das Ziel sollte so detailliert wie möglich beschrieben werden.
  • Budget: Welches Budget haben Sie für dieses Projekt eingeplant? Zusammen mit dem formulierten Ziel gibt das Budget die Wege vor, die wir zusammen in dem Projekt gehen können. Meist wollen Sie durch die Einführung von KI Kosten sparen oder höhere Umsätze erreichen. Das spielt für Höhe des Budgets die entscheidende Rolle.
  • Datenlage: Haben Sie Daten, die wir für das Training verwenden können? Wenn ja, welche und wieviele Daten sind das? Ist eine kontinuierliche Datenerfassung vorhanden, die während des Projekts genutzt werden kann, oder muss dafür erst die Grundlage geschaffen werden?

2. Evaluation

In diesem Schritt evaluieren und planen wir mit Ihnen gemeinsam die Umsetzung des Projekts. Das bedeutet im Einzelnen folgendes.

Begutachtung der Daten und weitere Datenplanung

Wir sichten von Ihnen bereitgestellte Trainingsdaten, z.B. gelabelte Bilder, und machen uns ein Bild davon, ob diese für das Training sinnvoll verwendet werden können. Da man für Deep Learning sehr viele Trainingsdaten benötigt, ist das ein entscheidender Punkt. In die Begutachtung der Daten fließt auch die Beurteilung der Qualität und Ausgewogenheit ein, denn davon ist abhängig wie gut ein KI-Modell lernt und korrekte Vorhersagen trifft.

Wenn von Ihnen keinerlei Daten zum Projektstart bereitgestellt werden können, wird zuerst ein separates Projekt notwendig, das nur dazu dient, Daten zu sammeln. Das bedeutet für Sie etwa je nach Anwendbarkeit den Einkauf von Datensets oder Labeling-Dienstleistungen.
Wir stehen Ihnen dabei beratend zur Seite.

Während der gesamten Dauer des Projekts werden immer wieder neue Daten benötigt, um die Qualität des Modells weiter zu verbessern. Daher müssen wir mit Ihnen gemeinsam planen, wie Sie fortlaufend diese Daten erheben, falsche Predictions des Modells erkennen und korrigieren, sodass Sie diese uns bereitstellen können. Die richtig erkannten Daten sowie die falsch erkannten und dann korrigierten Daten werden nämlich in das nächste Training einfließen.

Definition des Minimum Viable Product (MVP)

Wir definieren mit Ihnen zusammen, wie eine minimal funktionsfähige Version der KI aussehen kann. Die Grundfrage hierbei ist: Welche Komponenten oder Features sollten als Erstes in den Produktivbetrieb gehen, sodass Sie möglichst schnell einen Mehrwert aus
der KI ziehen?

Ein Vorteil dieser Herangehensweise ist, dass Sie den neuen KI-basierten Prozess in kleinem Maßstab testen können. Gleichzeitig können wir Verbesserungen schneller identifizieren. Zu einem späteren Zeitpunkt können Sie dann skalieren und weitere Features aufnehmen. Die schlagenden Argumente, mit einem MVP zu starten, sind jedoch die Kostenreduktion und Risikominimierung. Anstatt ein riesiges Projekt umzusetzen wird ein kleines Mehrwert schaffendes Projekt geschnürt und in der Realität getestet. So werden Fehlplanungen und
-entwicklungen vermieden, die viel Geld kosten.

Definition der Key Performance Indicators (KPI)

Key Performance Indicators sind für die objektive Qualitätsmessung der KI und des Business Impacts wichtig. Diese Zielmarken definieren, was das geplante System leisten soll, damit es erfolgreich ist. Key Performance Indicators können etwa sein:

  • Durchschnittliche Zeitersparnis des Prozesses durch Teilautomatisierung
  • Garantierte Antwortzeit bei maximalem Anfrageaufkommen pro Sekunde
  • Parallel mögliche Anfragen an die KI
  • Accuracy des Modells
  • Zeit von Fertigstellung bis zur Implementierung des KI Modells

Planung in Ihr Produktivsystem

Wir planen mit Ihnen die tiefe Integration in Ihr Produktivsystem. Dabei sind etwa folgende Fragen wichtig: Wie soll die KI in der bestehenden Softwareumgebung und im Arbeitsablauf genutzt werden? Was ist notwendig, um auf die KI zuzugreifen?

Mit dem Erstgespräch und der Evaluation ist nun das Fundament für das Projekt gelegt. In den Folgeschritten treiben wir die Entwicklung nun immer weiter voran. Die Schritte 3 bis 5 werden dabei solange wiederholt bis wir von der minimal funktionsfähigen
Produktversion, dem MVP, bis zum gewünschten Endprodukt gelangt sind.

3. Iteration

Wir trainieren den Algorithmus mit dem Großteil der verfügbaren Daten. Anschließend überprüfen wir die Performance des Modells mit ungesehenen Daten.

Wie lange das Training dauert ist abhängig von der Aufgabe. Man kann jedoch sagen, dass das Trainieren eines Deep Learning Modells für Bilder oder Videos komplexer und zeitaufwändiger ist als bei textbasierten maschinellen Lernaufgaben. Das liegt daran, dass wir tiefe Modelle (mit vielen Layern) verwenden und die verarbeiteten Datenmengen in der Regel sehr groß sind.

Das Trainieren des Modells ist je nach Projekt jedoch nur ein Bruchstück des ganzen Entwicklungsprozesses, den wir leisten. Oft ist es notwendig, dass wir einen eigenen Prozess aufbauen, in den das Modell eingebettet werden kann, wie z.B. einen Webservice.

4. Integration

Ist eine akzeptable Qualitätsstufe des Modells nach dem Training erreicht, liefern wir Ihnen eine erste Produktversion aus. Üblicherweise stellen wir Ihnen die Version als Docker Image mit API zur Verfügung. Sie beginnen dann mit der Integration in Ihr System und Ihre Workflows. Wir begleiten Sie dabei.

5. Feedback erfassen

Nachdem die Integration in den Produktivbetrieb erfolgt ist, ist es sehr wichtig, dass Sie aus der Nutzung Daten sammeln. Nur so können Sie beurteilen, ob die KI funktioniert wie Sie es sich vorgestellt haben und ob es in die richtige Richtung geht. Es geht also darum, zu erfassen was das Modell im Realbetrieb kann und was nicht. Diese Daten sammeln Sie und übermitteln sie an uns. Wir speisen diese dann in nächsten Trainingslauf ein.

Es ist dabei nicht ungewöhnlich, dass diese Datenerfassung im Realbetrieb eine gewisse Zeit in Anspruch nimmt. Das ist natürlich davon abhängig, in welchem Umfang Sie Daten erfassen. Bis zum Beginn der nächsten Iteration können so üblicherweise Wochen oder sogar Monate vergehen.

Die nächste Iteration

Um mit der nächsten Iteration eine signifikante Steigerung der Ergebnisqualität zu erreichen, kann es notwendig sein, dass Sie uns mehr Daten oder andere Daten zur Verfügung stellen, die aus dem Realbetrieb anfallen.

Eine nächste Iteration kann aber auch motiviert sein durch eine Veränderung in den Anforderungen, wenn etwa bei einem Klassifikationsmodell neue Kategorien erkannt werden müssen. Das aktuelle Modell kann für solche Veränderungen dann keine guten Vorhersagen treffen und muss erst mit entsprechenden neuen Daten trainiert werden.

Tipps für ein erfolgreiches KI Projekt

Ein entscheidender Knackpunkt für ein erfolgreiches KI Projekt ist das iterative Vorgehen und schrittweise Einführen eines KI-basierten Prozesses, mit dem die Qualität und Funktionsbreite der Entwicklung gesteigert wird.

Weiterhin muss man sich darüber klar sein, dass die Bereitstellung von Trainingsdaten kein statischer Ablauf ist. Es ist ein Kreislauf, in dem Sie als Kunde eine entscheidende Rolle einnehmen. Ein letzter wichtiger Punkt ist die Messbarkeit des Projekts. Denn nur wenn die Zielwerte während des Projekts gemessen werden, können Rückschritte oder Fortschritte gesehen werden und man kann schließlich am Ziel ankommen.

Möglich wurde dieser Artikel durch die großartige Zusammenarbeit mit pixolution, einem Unternehmen für AI Solutions im Bereich Computer Vision (Visuelle Bildsuche und individuelle KI Lösungen).

Deep Autoregressive Models

Deep Autoregressive Models

In this blog article, we will discuss about deep autoregressive generative models (AGM). Autoregressive models were originated from economics and social science literature on time-series data where obser- vations from the previous steps are used to predict the value at the current and at future time steps [SS05]. Autoregression models can be expressed as:

    \begin{equation*} x_{t+1}= \sum_i^t \alpha_i x_{t-i} + c_i, \end{equation*}

where the terms \alpha and c are constants to define the contributions of previous samples x_i for the future value prediction. In the other words, autoregressive deep generative models are directed and fully observed models where outcome of the data completely depends on the previous data points as shown in Figure 1.

Autoregressive directed graph.

Figure 1: Autoregressive directed graph.

Let’s consider x \sim X, where X is a set of images and each images is n-dimensional (n pixels). Then the prediction of new data pixel will be depending all the previously predicted pixels (Figure ?? shows the one row of pixels from an image). Referring to our last blog, deep generative models (DGMs) aim to learn the data distribution p_\theta(x) of the given training data and by following the chain rule of the probability, we can express it as:

(1)   \begin{equation*} p_\theta(x) = \prod_{i=1}^n p_\theta(x_i | x_1, x_2, \dots , x_{i-1}) \end{equation*}

The above equation modeling the data distribution explicitly based on the pixel conditionals, which are tractable (exact likelihood estimation). The right hand side of the above equation is a complex distribution and can be represented by any possible distribution of n random variables. On the other hand, these kind of representation can have exponential space complexity. Therefore, in autoregressive generative models (AGM), these conditionals are approximated/parameterized by neural networks.


As AGMs are based on tractable likelihood estimation, during the training process these methods maximize the likelihood of images over the given training data X and it can be expressed as:

(2)   \begin{equation*} \max_{\theta} \sum_{x\sim X} log \: p_\theta (x) = \max_{\theta} \sum_{x\sim X} \sum_{i=1}^n log \: p_\theta (x_i | x_1, x_2, \dots, x_{i-1}) \end{equation*}

The above expression is appearing because of the fact that DGMs try to minimize the distance between the distribution of the training data and the distribution of the generated data (please refer to our last blog). The distance between two distribution can be computed using KL-divergence:

(3)   \begin{equation*} \min_{\theta} d_{KL}(p_d (x),p_\theta (x)) = log\: p_d(x) - log \: p_\theta(x) \end{equation*}

In the above equation the term p_d(x) does not depend on \theta, therefore, whole equation can be shortened to Equation 2, which represents the MLE (maximum likelihood estimation) objective to learn the model parameter \theta by maximizing the log likelihood of the training images X. From implementation point of view, the MLE objective can be optimized using the variations of stochastic gradient (ADAM, RMSProp, etc.) on mini-batches.

Network Architectures

As we are discussing deep generative models, here, we would like to discuss the deep aspect of AGMs. The parameterization of the conditionals mentioned in Equation 1 can be realized by different kind of network architectures. In the literature, several network architectures are proposed to increase their receptive fields and memory, allowing more complex distributions to be learned. Here, we are mentioning a couple of well known architectures, which are widely used in deep AGMs:

  1. Fully-visible sigmoid belief network (FVSBN): FVSBN is the simplest network without any hidden units and it is a linear combination of the input elements followed by a sigmoid function to keep output between 0 and 1. The positive aspects of this network is simple design and the total number of parameters in the model is quadratic which is much smaller compared to exponential [GHCC15].
  2. Neural autoregressive density estimator (NADE): To increase the effectiveness of FVSBN, the simplest idea would be to use one hidden layer neural network instead of logistic regression. NADE is an alternate MLP-based parameterization and more effective compared to FVSBN [LM11].
  3. Masked autoencoder density distribution (MADE): Here, the standard autoencoder neural networks are modified such that it works as an efficient generative models. MADE masks the parameters to follow the autoregressive property, where the current sample is reconstructed using previous samples in a given ordering [GGML15].
  4. PixelRNN/PixelCNN: These architecture are introducced by Google Deepmind in 2016 and utilizing the sequential property of the AGMs with recurrent and convolutional neural networks.
Different autoregressive architectures

Figure 2: Different autoregressive architectures (image source from [LM11]).

Results using different architectures

Results using different architectures (images source

It uses two different RNN architectures (Unidirectional LSTM and Bidirectional LSTM) to generate pixels horizontally and horizontally-vertically respectively. Furthermore, it ulizes residual connection to speed up the convergence and masked convolution to condition the different channels of images. PixelCNN applies several convolutional layers to preserve spatial resolution and increase the receptive fields. Furthermore, masking is applied to use only the previous pixels. PixelCNN is faster in training compared to PixelRNN. However, the outcome quality is better with PixelRNN [vdOKK16].


In this blog article, we discussed about deep autoregressive models in details with the mathematical foundation. Furthermore, we discussed about the training procedure including the summary of different network architectures. We did not discuss network architectures in details, we would continue the discussion of PixelCNN and its variations in upcoming blogs.


[GGML15] Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: masked autoencoder for distribution estimation. CoRR, abs/1502.03509, 2015.

[GHCC15] Zhe Gan, Ricardo Henao, David Carlson, and Lawrence Carin. Learning Deep Sigmoid Belief Networks with Data Augmentation. In Guy Lebanon and S. V. N. Vishwanathan, editors, Proceedings of the Eighteenth International Conference on Artificial Intelligence
and Statistics, volume 38 of Proceedings of Machine Learning Research, pages 268–276, San Diego, California, USA, 09–12 May 2015. PMLR.

[LM11] Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In Geoffrey Gordon, David Dunson, and Miroslav Dudík, editors, Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 of Proceedings of Machine Learning Research, pages 29–37, Fort Lauderdale, FL, USA, 11–13 Apr 2011.

[SS05] Robert H. Shumway and David S. Stoffer. Time Series Analysis and Its Applications (Springer Texts in Statistics). Springer-Verlag, Berlin, Heidelberg, 2005.

[vdOKK16] A ̈aron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural
networks. CoRR, abs/1601.06759, 2016

What is Portfolio Risk Management in Python?

Data science is a crucial industry, with multiple processes today relying on it. One of its more helpful and intriguing applications is in investing, where it helps investors make more informed decisions. Practices like portfolio management in Python help take the guesswork out of this notoriously risky undertaking.

Investing is a complicated science, making it hard to do well. Some estimates hold that as much as 90% of people lose money in stocks. While stock trading will always involve some risk, Python-based portfolio management can help.

What Is Portfolio Management in Python?

Portfolio management is the process of planning, making and overseeing investments to meet your long-term investment goals. Portfolio management in Python uses data science to analyze risks and rewards to make the best investment decisions.

Since the future is uncertain, buying stocks is inherently risky, but some assets are riskier than others. For example, since many companies are trying to reach carbon neutrality by 2050, investing in sustainable technologies is a fairly sound strategy. However, that doesn’t guarantee that every eco-friendly startup will succeed, so investors need to consider more factors.

Some data scientists have found that you can use Python to understand these factors better. By plugging various figures into a Python equation, investors can chart potential risks and returns to find the best investments.

How Does Python Portfolio Management Work?

Portfolio risk management in Python operates on a principle called Modern Portfolio Theory (MPT). MPT helps investors find an optimal mix of high-risk, high-return investments and low-risk, low-return ones based on their risk tolerance. Investors can either look for the highest returns at a certain risk level or look for the lowest risk to get a certain return.

To apply this in Python, data scientists create one list for portfolio returns, one for risk and one for weights, or how much each investment accounts for the overall portfolio. They then randomly generate weight for the assets, then normalize it to sum to a value of one.

Data scientists then calculate the risks and returns for each asset and plug them into the different randomly generated weights. This will produce a list of various scenarios, showing how much overall risk and reward each portfolio would have.

Investors can then look at this list to see how much of each asset they should include in their portfolio. They can either use the mix that produces the greatest return or the one with the lowest risk.

Why Does It Matter?

Using Python for portfolio risk management helps remove a lot of the guesswork from investing. Running these calculations gives investors multiple scenarios to choose from, helping them find the best portfolio strategy for their needs and goals.

This presents a promising opportunity for data scientists. Data analytics are quickly becoming an essential part of the stock market. Algorithmic trading, which applies data and AI to MPT, already accounts for 60 to 73% of all U.S. equity trading. Portfolio management in Python could help more data scientists capitalize on this trend.

This practice is a relatively straightforward way to apply data science to stock trading. Data scientists that can make the most of that opportunity stand to make a name for themselves in investing circles.

Python Portfolio Management Can Maximize Returns

In the past, stock trading was almost akin to gambling, involving huge amounts of risk. While portfolio management in Python doesn’t remove volatility from the stock market, it helps put it in perspective. Investors can then make safer, more informed decisions to meet their investing goals.

Python-based portfolio management stands as a natural intersection between data science and stock trading. As a result, it can help both data scientists and investors achieve new success.

Vorstellung des Verbundforschungsprojekts “What can AI do for me?”

Dieser Artikel ist eine Zusammenfassung der Ergebnisse einer Studie namens “What can AI do for me?” ( Ansprechpartnerin für dieses Projekt ist Frau Carina Weber, Mitarbeiterin der Hochschule der Medien in Stuttgart.

Hintergrund zur Studie: Zu dem Thema Anwendung von Künstliche Intelligenz und ihrem Potenzial für die Wertschöpfung von Unternehmen gibt es bereits einige wenige Studien. Die wenigen Forschungsarbeiten stellen positive Auswirkungen, wie Produktoptimierung, Kosteneinsparung durch Optimierung des Ressourcenmanagements, Steigerung der allgemein Unternehmensperformance, etc. fest. Allerdings bleibt unerforscht welchen individuellen Beitrag spezifische Anwendungsfälle leisten. Dieses Wissen wird jedoch für strategische Entscheidungen bezüglich der Implementierung von AI benötigt, um beispielsweise den ROI von AI-Projekten schätzen zu können. Dazu soll die vorliegende Studie Einsicht bringen.

Darüberhinaus wurden die Ergebnisse genutzt um im Rahmen des Verbundforschungsprojekts What Can AI Do For Me? eine AI-basierte Matching-Plattform zu entwickeln. Eine bis jetzt einzigartige Anwendung, mittels derer Unternehmen individuelle AI-Anwendungsfälle mit ihren jeweiligen Potenzialen kennenlernen und sich direkt mit Lösungsanbietern verknüpfen lassen können.

Beispiele: Praktische Anwendung von AI – Mit welchen Herausforderungen sehen sich Unternehmen konfrontiert?

Schon heute stellt Artificial Intelligence, folgend abgekürzt mit AI, im unternehmerischen

Sinne eine Schlüsseltechnologie dar. Es stellt sich jedoch die Frage, inwieweit sich die Technologien rund um AI tatsächlich auf die essentiellen Unternehmensziele auswirken und mit welchen Hindernissen sich die Unternehmen bei der Implementierung konfrontiert sehen.

In der AI Value Creation Studie des Forschungsprojekts “What can AI do for me” ist man mit Unterstützung von Expertinnen und Experten, sowohl auf Anwenderseite, als auch auf der von Nutzerinnen und Nutzer, dieser Fragestellung, durch eine qualitative und quantitative Forschung nachgegangen.

Unsicher beim Einsatz von AI? Die Studie bietet Orientierungshilfe

Das Institute of Applied Artificial Intelligence (IAAI) der Hochschule der Medien entwickelt im Rahmen des oben genannten Verbundforschungsprojekts zusammen mit der thingsTHINKING GmbH und der KENBUN IT AG eine AI-basierte Matching-Plattform, mittels derer Unternehmen geeignete Anwendungsmöglichkeiten und Lösungsunternehmen finden können. Gefördert wurde das Projekt im Jahr 2021 über den KI-Innovationswettbewerb des Ministeriums für Wirtschaft, Arbeit und Tourismus Baden-Württemberg und erhielt zusätzliche Unterstützung von bekannten AI-Initiativen und Verbänden. So konnte am 19. Oktober die Inbetriebnahme der Beta-Version erfolgreich gestartet werden. Sie steht seitdem unter der Domain kostenfrei zur Verfügung.

Die Basis der Annahmen der Matching-Plattform bilden die Ergebnisse der AI Value Creation Studie des IAAI der Hochschule der Medien. Im Verlauf der qualitativen Forschung konnten über 90 verschiedene AI Use Cases aus der Unternehmenspraxis in über 40 Interviews mit Expertinnen und -experten vielfältigster Branchen identifiziert werden. Die erhobenen Use Cases wurden in insgesamt 19 Use Case Cluster strukturiert, um eine bessere Vergleichbarkeit zu schaffen und gleichzeitig vielfältige Anwendungsmöglichkeiten aufzuzeigen.

Es wird eine Orientierungshilfe für Unternehmen geschaffen, über die sie einen Überblick erlangen können, in welchen Unternehmensfunktionen AI bereits erfolgreich eingesetzt wird.

Des Weiteren sollen durch die Studie Potenziale von AI in Bezug auf die Wertschöpfung, im Sinne einer möglichen Umsatz-, Unternehmenswertsteigerung sowie Kostensenkung, erhoben und Hindernisse bei der Realisierung von AI Use Cases erkannt werden. Zuletzt sollen Unternehmen dazu befähigt werden Stellschrauben zu identifizieren, an welchen sie ansetzen müssen, um AI erfolgreich im Unternehmen einzusetzen.

Im Rahmen der erhobenen Studie wurde einerseits eine Dominanz der AI Use Cases im Bereich der Produktion und Supply Chain, Marketing und Sales sowie im Kundenservice deutlich. Andererseits konnten vielzählige Use Cases ermittelt werden, die crossfunktional in Unternehmen eingesetzt werden können und somit wiederkehrende Tätigkeiten, wie AI-gestützte Recherche in Datenbanken oder Sachbearbeitung von Dokumenten, in Unternehmen unterstützen.

Variierendes Wertschöpfungspotenzial je nach Einsatzbereich und Aufgabe

Gerade bei Use Cases mit AI-Anwendungen, die über verschiedeneUnternehmensfunktionen hinweg eingesetzt werden können, ist die Einschätzung des Wertschöpfungspotenzials abhängig von der individuellen Aufgabe und dem Anwendungsbereich und demnach sehr divers.

Über alle erhobenen Use Cases hinweg tendieren die befragten Personen dazu das Wertschöpfungspotenzial zur Kostenreduktion am höchsten einzuschätzen. Dieses Phänomen kann dadurch erklärt werden, dass ineffiziente Prozesse schnell zu höheren Kosten führen, bei einer beschleunigten, zuverlässigeren Ausführung durch AI das Potenzial zur Kostenersparnis schnell ersichtlich werden kann. Dadurch wurde dieses Wertschöpfungspotenzial im Vergleich zu Umsatz- und Unternehmenswertsteigerung auch häufiger von Expertinnen und Experten identifiziert. Zusätzlich zu diesen Erkenntnissen wurden in Interviews weitere Aspekte bzw. Ziele des

AI-Einsatzes in den Unternehmen abgefragt, die sich abseits schon genannten Wertschöpfungspotenziale indirekt auf die Wertschöpfung und den Unternehmenserfolg auswirken. So wurden neben Prozessoptimierung, die Steigerung der ökologischen und ökonomischen Nachhaltigkeit, die Verbesserung des Unternehmensimages und eine Steigerung der Unternehmensattraktivität genannt.

Fehlende Daten, fehlendes Personal – die Hindernisse bei der Implementierung

In der qualitativen Studie wurden neben den Potenzialen von AI auch Hindernisse und Herausforderungen. Durch eine genaue Systematisierung und Analyse wurde deutlich: der Mangel an Daten, personellen und finanziellen Ressourcen und das fehlendes Mindset machen den Unternehmen zu schaffen. Um diese Ergebnisse besser beurteilen und einschätzen zu können wurden Branchenexpertinnen und -experten gebeten, die ermittelten Herausforderungen im Rahmen einer quantitativen Studie zu bewerten. Die Stichprobe besteht aus Mitarbeiterinnen und Mitarbeiter in beratender Funktion bei AI-Projekte, Managerinnen und Manager mit Entscheidungsfunktion auf diesem Gebiet sowie Unternehmensberaterinnen  und -berater aus Beratungsfirmen mit Fokus auf AI-Projekten.

Sehr deutlich wurde hierbei der allgegenwärtige Mangel an Fachpersonal, der von weit mehr als der Hälfte der Befragten angegeben wurde. Zudem ist die gegebenen Datenqualität oft nur unzureichend und es fehlt an AI-Strategien, was sehr große Hindernisse angesehen wurden. Im Vergleich hierzu waren Hindernisse wie ein mangelnder Reifegrad der AI-Technologien und offene Rechtsfragen nur von etwas mehr als einem Drittel der Befragten angegeben worden. Was natürlich zum einen deutlich macht, dass zwar verschiedene Herausforderungen bei der AI-Implementierung gibt, es aber oft in den Händen der Unternehmen liegt inwieweit diese überwunden werden.

Weiterführende Informationen zum Forschungsbericht und dem Projekt

Weitere Ergebnisse und Informationen zur Forschungsmethode können dem Forschungsbericht der Autoren Prof. Dr. Jürgen Seitz, Katharina Willbold, Robin Haiber und Alicia Krafft entnommen werden. Dieser kann vollständig kostenlos unter eingesehen werden. Weiterhin steht die AI-basierte Matching-Plattform WhatCanAIDoForMe? des IAAI der Hochschule der Medien, der thingsTHINKING Gmbh und der KENBUN IT AG kostenfrei zur Anwendung bereit.

Hier werden Unternehmen ausgehend von einer Beschreibung zur Problemstellung ihres Business Cases über ein semantisches Matching passende AI-Anwendungsfälle vorgeschlagen. Darüber hinaus wird ein numerisches Wertschöpfungspotenzial aus Basis einer Expertinnen-/ Expertenmeinung angezeigt. Dieses kann als ein erster Indikator für eine Bewertung des AI-Vorhabens herangezogen werden.

Unter der Domain kann die Plattform aufgerufen werden.


Jürgen Seitz
Dr. Jürgen Seitz ist einer der führenden Professoren im Bereich Digitalisierung in Deutschland. Als Mitbegründer, Geschäftsführer und Beirat hat er geholfen, mehrere erfolgreiche digitale Unternehmen aufzubauen und zu skalieren. Seine beruflichen Stationen umfassten u.A. Microsoft, WEB.DE und die United Internet Gruppe (1&1). Heute forscht und lehrt er an der Hochschule der Medien in Stuttgart in den Bereichen Digital Marketing und Digital Business. Er ist außerdem Gründungsprofessor am Institute for Applied Artificial Intelligence (IAAI), Herausgeber der Digital Insights Studienreihe und engagiert sich für die Digitalisierung von NGOs.
Alicia Krafft
Alicia Krafft, Studentin an der Hochschule der Medien in Stuttgart, absolviert derzeit ihr Masterstudium in Unternehmenskommunikation mit den Schwerpunkten Digitale Medien und Marketing sowie Web Analytics. In den letzten Jahren half sie digitale Kommunikationsstrategien für diverse Unternehmen zu entwickeln und umzusetzen, u.a. für die ARENA2036, ein Forschungscampus der Universität Stuttgart, und zuletzt für das Forschungsteam rund um Dr. Jürgen Seitz.

Graphical understanding of dynamic programming and the Bellman equation: taking a typical approach at first

This is the second article of the series My elaborate study notes on reinforcement learning.

*I must admit I could not fully explain how I tried visualizing ideas of Bellman equations in this article. I highly recommend you to also take brief at the second section of the third article. (A comment added on 13/3/2022)

1, Before getting down on business

As the title of this article suggests, this article is going to be mainly about the Bellman equation and dynamic programming (DP), which are to be honest very typical and ordinary topics. One typical way of explaining DP in contexts of reinforcement learning (RL) would be explaining the Bellman equation, value iteration, and policy iteration, in this order. If you would like to merely follow pseudocode of them and implement them, to be honest that is not a big deal. However even though I have studied RL only for some weeks, I got a feeling that these algorithms, especially policy iteration are more than just single algorithms. In order not to miss the points of DP, rather than typically explaining value iteration and policy iteration, I would like to take a different approach. Eventually I am going to introduce DP in RL as a combination of the following key terms: the Bellman operator, the fixed point of a policy, policy evaluation, policy improvement, and existence of the optimal policy. But first, in this article I would like to cover basic and typical topics of DP in RL.

Many machine learning algorithms which use supervised/unsupervised learning more or less share the same ideas. You design a model and a loss function and input samples from data, and you adjust parameters of the model so that the loss function decreases. And you usually use optimization techniques like stochastic gradient descent (SGD) or ones derived from SGD. Actually feature engineering is needed to extract more meaningful information from raw data. Or especially in this third AI boom, the models are getting more and more complex, and I would say the efforts of feature engineering was just replaced by those of designing neural networks. But still, once you have the whole picture of supervised/unsupervised learning, you would soon realize other various algorithms is just a matter of replacing each component of the workflow. However reinforcement learning has been another framework of training machine learning models. Richard E. Bellman’s research on DP in 1950s is said to have laid a foundation for RL. RL also showed great progress thanks to development of deep neural networks (DNN), but still you have to keep it in mind that RL and supervised/unsupervised learning are basically different frameworks. DNN are just introduced in RL frameworks to enable richer expression of each component of RL. And especially when RL is executed in a higher level environment, for example screens of video games or phases of board games, DNN are needed to process each state of the environment. Thus first of all I think it is urgent to see ideas unique to RL in order to effectively learn RL. In the last article I said RL is an algorithm to enable planning by trial and error in an environment, when the model of the environment is not known. And DP is a major way of solving planning problems. But in this article and the next article, I am mainly going to focus on a different aspect of RL: interactions of policies and values.

According to a famous Japanese textbook on RL named “Machine Learning Professional Series: Reinforcement Learning,” most study materials on RL lack explanations on mathematical foundations of RL, including the book by Sutton and Barto. That is why many people who have studied machine learning often find it hard to get RL formulations at the beginning. The book also points out that you need to refer to other bulky books on Markov decision process or dynamic programming to really understand the core ideas behind algorithms introduced in RL textbooks. And I got an impression most of study materials on RL get away with the important ideas on DP with only introducing value iteration and policy iteration algorithms. But my opinion is we should pay more attention on policy iteration. And actually important RL algorithms like Q learning, SARSA, or actor critic methods show some analogies to policy iteration. Also the book by Sutton and Barto also briefly mentions “Almost all reinforcement learning methods are well described as GPI (generalized policy iteration). That is, all have identifiable policies and value functions, with the policy always being improved with respect to the value function and the value function always being driven toward the value function for the policy, as suggested by the diagram to the right side.

Even though I arrogantly, as a beginner in this field, emphasized “simplicity” of RL in the last article, in this article I am conversely going to emphasize the “profoundness” of DP over two articles. But I do not want to cover all the exhaustive mathematical derivations for dynamic programming, which would let many readers feel reluctant to study RL. I tried as hard as possible to visualize the ideas in DP in simple and intuitive ways, as far as I could understand. And as the title of this article series shows, this article is also a study note for me. Any corrections or advice would be appreciated via email or comment pots below.

2, Taking a look at what DP is like

In the last article, I said that planning or RL is a problem of finding an optimal policy \pi(a|s) for choosing which actions to take depending on where you are. Also in the last article I displayed flows of blue arrows for navigating a robot as intuitive examples of optimal policies in planning or RL problems. But you cannot directly calculate those policies. Policies have to be evaluated in the long run so that they maximize returns, the sum of upcoming rewards. Then in order to calculate a policy p(a|s), you need to calculate a value functions v_{\pi}(s). v_{\pi}(s) is a function of how good it is to be in a given state s, under a policy \pi. That means it is likely you get higher return starting from s, when v_{\pi}(s) is high. As illustrated in the figure below, values and policies, which are two major elements of RL, are updated interactively until they converge to an optimal value or an optimal policy. The optimal policy and the optimal value are denoted as v_{\ast} and \pi_{\ast} respectively.

Dynamic programming (DP) is a family of algorithms which is effective for calculating the optimal value v_{\ast} and the optimal policy \pi_{\ast} when the complete model of the environment is given. Whether in my articles or not, the rest of discussions on RL are more or less based on DP. RL can be viewed as a method of achieving the same effects as DP when the model of the environment is not known. And I would say the effects of imitating DP are often referred to as trial and errors in many simplified explanations on RL. If you have studied some basics of computer science, I am quite sure you have encountered DP problems. With DP, in many problems on textbooks you find optimal paths of a graph from a start to a goal, through which you can maximizes the sum of scores of edges you pass. You might remember you could solve those problems in recursive ways, but I think many people have just learnt very limited cases of DP. For the time being I would like you to forget such DP you might have learned and comprehend it as something you newly start learning in the context of RL.

*As a more advances application of DP, you might have learned string matching. You can calculated how close two strings of characters are with DP using string matching.

The way of calculating v_{\pi}(s) and \pi(a|s) with DP can be roughly classified to two types, policy-based and value-based. Especially in the contexts of DP, the policy-based one is called policy iteration, and the values-based one is called value iteration. The biggest difference between them is, in short, policy iteration updates a policy every times step, but value iteration does it only at the last time step. I said you alternate between updating v_{\pi}(s) and \pi(a|s), but in fact that is only true of policy iteration. Value iteration updates a value function v(s). Before formulating these algorithms, I think it will be effective to take a look at how values and policies are actually updated in a very simple case. I would like to introduce a very good tool for visualizing value/policy iteration. You can customize a grid map and place either of “Treasure,” “Danger,” and “Block.” You can choose probability of transition and either of settings, “Policy Iteration” or “Values Iteration.” Let me take an example of conducting DP on a gird map like below. Whichever of “Policy Iteration” or “Values Iteration” you choose, you would get numbers like below. Each number in each cell is the value of each state, and you can see that when you are on states with high values, you are more likely to reach the “treasure” and avoid “dangers.” But I bet this chart does not make any sense if you have not learned RL yet. I prepared some code for visualizing the process of DP on this simulator. The code is available in this link.

*In the book by Sutton and Barto, when RL/DP is discussed at an implementation level, the estimated values of v_{\pi}(s) or v_{\ast}(s) can be denoted as an array V or V_t. But I would like you take it easy while reading my articles. I will repeatedly mentions differences of notations when that matters.

*Remember that at the beginning of studying RL, only super easy cases are considered, so a V is usually just a NumPy array or an Excel sheet.

*The chart above might be also misleading since there is something like a robot at the left bottom corner, which might be an agent. But the agent does not actually move around the environment in planning problems because it has a perfect model of the environment in the head.

The visualization I prepared is based on the implementation of the simulator, so they would give the same outputs. When you run policy iteration in the map, the values and polices are updated as follows. The arrow in each cell is the policy in the state. At each time step the arrows is calculated in a greedy way, and each arrow at each state shows the direction in which the agent is likely to get the highest reward. After 3 iterations, the policies and values converge, and with the policies you can navigate yourself to the “Treasure,” avoiding “Dangers.”

*I am not sure why policies are incorrect at the most left side of the grid map. I might need some modification of code.

You can also update values without modifying policies as the chart below. In this case only the values of cells are updated. This is value-iteration, and after this iteration converges, if you transit to an adjacent cell with the highest value at each cell, you can also navigate yourself to the “treasure,” avoiding “dangers.”

I would like to start formulating DP little by little,based on the notations used in the RL book by Sutton. From now on, I would take an example of the 5 \times 6 grid map which I visualized above. In this case each cell is numbered from 0 to 29 as the figure below. But the cell 7, 13, 14 are removed from the map. In this case \mathcal{S} = {0, 1, 2, 3, 4, 6, 8, 9, 10, 11, 12, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29}, and \mathcal{A} = \{\uparrow, \rightarrow, \downarrow, \leftarrow \}. When you pass s=8, you get a reward r_{treasure}=1, and when you pass the states s=15 or s=19, you get a reward r_{danger}=-1. Also, the agent is encouraged to reach the goal as soon as possible, thus the agent gets a regular reward of r_{regular} = - 0.04 every time step.

In the last section, I mentioned that the purpose of RL is to find the optimal policy which maximizes a return, the sum of upcoming reward R_t. A return is calculated as follows.

R_{t+1} + R_{t+2} +  R_{t+3} + \cdots + R_T

In RL a return is estimated in probabilistic ways, that is, an expectation of the return given a state S_t = s needs to be considered. And this is the value of the state. Thus the value of a state S_t = s is calculated as follows.

\mathbb{E}_{\pi}\bigl[R_{t+1} + R_{t+2} +  R_{t+3} + \cdots + R_T | S_t = s \bigr]

In order to roughly understand how this expectation is calculated let’s take an example of the 5 \times 6 grid map above. When the current state of an agent is s=10, it can take numerous patterns of actions. For example (a) 10 - 9 - 8 - 2 , (b) 10-16-15-21-20-19, (c) 10-11-17-23-29-\cdots. The rewards after each behavior is calculated as follows.

  • If you take a you take the course (a) 10 - 9 - 8 - 2, you get a reward of r_a = -0.04 -0.04 + 1 -0.04 in total. The probability of taking a course of a) is p_a = \pi(A_t = \leftarrow | S_t = 10) \cdot p(S_{t+1} = 9 |S_t = 10, A_t = \leftarrow ) \cdot \pi(A_{t+1} = \leftarrow | S_{t+1} = 9) \cdot p(S_{t+2} = 8 |S_{t+1} = 9, A_{t+1} = \leftarrow ) \cdot \pi(A_{t+2} = \uparrow | S_{t+2} = 8) \cdot p(S_{t+3} = 2 | S_{t+2} = 8, A_{t+2} = \uparrow )
  • Just like the case of (a), the reward after taking the course (b) is r_b = - 0.04 -0.04 -1 -0.04 -0.04 -0.04 -1. The probability of taking the action can be calculated in the same way as p_b = \pi(A_t = \downarrow | S_t = 10) \cdot p(S_{t+1} = 16 |S_t = 10, A_t = \downarrow ) \cdots \pi(A_{t+4} = \leftarrow | S_{t+4} = 20) \cdot p(S_{t+5} = 19 |S_{t+4} = 20, A_{t+4} = \leftarrow ).
  • The rewards and the probability of the case (c) cannot be calculated because future behaviors of the agent is not confirmed.

Assume that (a) and (b) are the only possible cases starting from s, under the policy \pi, then the the value of s=10 can be calculated as follows as a probabilistic sum of rewards of each behavior (a) and (b).

\mathbb{E}_{\pi}\bigl[R_{t+1} + R_{t+2} +  R_{t+3} + \cdots + R_T | S_t = s \bigr] = r_a \cdot p_a + r_b \cdot p_b

But obviously this is not how values of states are calculated in general. Starting from a state a state s=10, not only (a) and (b), but also numerous other behaviors of agents can be considered. Or rather, it is almost impossible to consider all the combinations of actions, transition, and next states. In practice it is quite difficult to calculate a sequence of upcoming rewards R_{t+1}, \gamma R_{t+2}, R_{t+3} \cdots,and it is virtually equal to considering all the possible future cases.A very important formula named the Bellman equation effectively formulate that.

3, The Bellman equation and convergence of value functions

*I must admit I could not fully explain how I tried visualizing ideas of Bellman equations in this article. It might be better to also take brief at the second section of the third article. (A comment added on 3/3/2022)

The Bellman equation enables estimating values of states considering future countless possibilities with the following two ideas.

  1.  Returns are calculated recursively.
  2.  Returns are calculated in probabilistic ways.

First of all, I have to emphasize that a discounted return is usually used rather than a normal return, and a discounted one is defined as below

G_t \doteq R_{t+1} + \gamma R_{t+2} + \gamma ^2 R_{t+3} + \cdots + \gamma ^ {T-t-1} R_T = \sum_{k=0}^{T-t-1}{\gamma ^{k}R_{t+k+1}}

, where \gamma \in (0, 1] is a discount rate. (1)As the first point above, the discounted return can be calculated recursively as follows: G_t = R_{t + 1} + \gamma R_{t + 2} + \gamma ^2 R_{t + 2} + \gamma ^3 R_{t + 3} + \cdots = R_{t + 1} + \gamma (R_{t + 2} + \gamma R_{t + 2} + \gamma ^2 R_{t + 3} + \cdots ) = R_{t + 1} + \gamma G_{t+1}. You can postpone calculation of future rewards corresponding to G_{t+1} this way. This might sound obvious, but this small trick is crucial for defining defining value functions or making update rules of them. (2)The second point might be confusing to some people, but it is the most important in this section. We took a look at a very simplified case of calculating the expectation in the last section, but let’s see how a value function v_{\pi}(s) is defined in the first place.

v_{\pi}(s) \doteq \mathbb{E}_{\pi}\bigl[G_t | S_t = s \bigr]

This equation means that the value of a state s is a probabilistic sum of all possible rewards taken in the future following a policy \pi. That is, v_{\pi}(s) is an expectation of the return, starting from the state s. The definition of a values v_{\pi}(s) is written down as follows, and this is what \mathbb{E}_{\pi} means.

v_{\pi} (s)= \sum_{a}{\pi(a|s) \sum_{s', r}{p(s', r|s, a)\bigl[r + \gamma v_{\pi}(s')\bigr]}}

This is called Bellman equation, and it is no exaggeration to say this is the foundation of many of upcoming DP or RL ideas. Bellman equation can be also written as \sum_{s', r, a}{\pi(a|s) p(s', r|s, a)\bigl[r + \gamma v_{\pi}(s')\bigr]}. It can be comprehended this way: in Bellman equation you calculate a probabilistic sum of r +v_{\pi}(s'), considering all the possible actions of the agent in the time step. r +v_{\pi}(s') is a sum of the values of the next state s' and a reward r, which you get when you transit to the state s' from s. The probability of getting a reward r after moving from the state s to s', taking an action a is \pi(a|s) p(s', r|s, a). Hence the right side of Bellman equation above means the sum of \pi(a|s) p(s', r|s, a)\bigl[r + \gamma v_{\pi}(s')\bigr], over all possible combinations of s', r, and a.

*I would not say this equation is obvious, and please let me explain a proof of this equation later.

The following figures are based on backup diagrams introduced in the book by Sutton and Barto. As we have just seen, Bellman expectation equation calculates a probabilistic summation of r + v(s'). In order to calculate the expectation, you have to consider all the combinations of s', r, and a. The backup diagram at the left side below shows the idea as a decision-tree-like graph, and strength of color of each arrow is the probability of taking the path.

The Bellman equation I have just introduced is called Bellman expectation equation to be exact. Like the backup diagram at the right side, there is another type of Bellman equation where you consider only the most possible path. Bellman optimality equation is defined as follows.

v_{\ast}(s) \doteq \max_{a} \sum_{s', r}{p(s', r|s, a)\bigl[r + \gamma v_{\ast}(s')\bigr]}

I would like you to pay attention again to the fact that in definitions of Bellman expectation/optimality equations, v_{\pi}(s)/v_{\ast}(s) is defined recursively with v_{\pi}(s)/v_{\ast}(s). You might have thought how to calculate v_{\pi}(s)/v_{\ast}(s) is the problem in the first place.

As I implied in the first section of this article, ideas behind how to calculate these v_{\pi}(s) and v_{\ast}(s) should be discussed more precisely. Especially how to calculate v_{\pi}(s) is a well discussed topic in RL, including the cases where data is sampled from an unknown environment model. In this article we are discussing planning problems, where a model an environment is known. In planning problems, that is DP problems where all the probabilities of transition p(s', r | s, a) are known, a major way of calculating v_{\pi}(s) is iterative policy evaluation. With iterative policy evaluation a sequence of value functions (v_0(s), v_1(s), \dots , v_{k-1}(s), v_{k}(s)) converges to v_{\pi}(s) with the following recurrence relation

v_{k+1}(s) =\sum_{a}{\pi(a|s)\sum_{s', r}{p(s', r | s, a) [r + \gamma v_k (s')]}}.

Once v_{k}(s) converges to v_{\pi}(s), finally the equation of the definition of v_{\pi}(s) holds as follows.

v_{\pi}(s) =\sum_{a}{\pi(a|s)\sum_{s', r}{p(s', r | s, a) [r + \gamma v_{\pi} (s')]}}.

The convergence to v_{\pi}(s) is like the graph below. If you already know how to calculate forward propagation of a neural network, this should not be that hard to understand. You just expand recurrent relation of v_{k}(s) and v_{k+1}(s) from the initial value at k=0 to the converged state at k=K. But you have to be careful abut the directions of the arrows in purple. If you correspond the backup diagrams of the Bellman equation with the graphs below, the purple arrows point to the reverse side to the direction where the graphs extend. This process of converging an arbitrarily initialized v_0(s) to v_{\pi}(s) is called policy evaluation.

*\mathcal{S}, \mathcal{A} are a set of states and actions respectively. Thus |\mathcal{S}|, the size of  \mathcal{S} is the number of white nodes in each layer, and |\mathcal{S}| the number of black nodes.

The same is true of the process of calculating an optimal value function v_{\ast}. With the following recurrence relation

v_{k+1}(s) =\max_a\sum_{s', r}{p(s', r | s, a) [r + \gamma v_k (s')]}

(v_0(s), v_1(s), \dots , v_{k-1}(s), v_{k}(s)) converges to an optimal value function v_{\ast}(s). The graph below visualized the idea of convergence.

4, Pseudocode of policy iteration and value iteration

I prepared pseudocode of each algorithm based on the book by Sutton and Barto. These would be one the most typical DP algorithms you would encounter while studying RL, and if you just want to implement RL by yourself, these pseudocode would enough. Or rather these would be preferable to other more general and abstract pseudocode. But I would like to avoid explaining these pseudocode precisely because I think we need to be more conscious about more general ideas behind DP, which I am going to explain in the next article. I will cover only the important points of these pseudocode, and I would like to introduce some implementation of the algorithms in the latter part of next article. I think you should briefly read this section and come back to this section section or other study materials after reading the next article. In case you want to check the algorithms precisely, you could check the pseudocode I made with LaTeX in this link.

The biggest difference of policy iteration and value iteration is the timings of updating a policy. In policy iteration, a value function v(s) and \pi(a|s) are arbitrarily initialized. (1)The first process is policy evaluation. The policy \pi(a|s) is fixed, and the value function v(s) approximately converge to v_{\pi}(s), which is a value function on the policy \pi. This is conducted by the iterative calculation with the reccurence relation introduced in the last section.(2) The second process is policy improvement. Based on the calculated value function v_{\pi}(s), the new policy \pi(a|s) is updated as below.

\pi(a|s) \gets\text{argmax}_a {r + \sum_{s', r}{p(s', r|s, a)[r + \gamma V(s')]}}, \quad \forall s\in \mathcal{S}

The meaning of this update rule of a policy is quite simple: \pi(a|s) is updated in a greedy way with an action a such that r + \sum_{s', r}{p(s', r|s, a)[r + \gamma V(s')]} is maximized. And when the policy \pi(a|s) is not updated anymore, the policy has converged to the optimal one. At least I would like you to keep it in mind that a while loop of itrative calculation of v_{\pi}(s) is nested in another while loop. The outer loop continues till the policy is not updated anymore.

On the other hand in value iteration, there is mainly only one loop of updating  v_{k}(s), which converge to v_{\ast}(s). And the output policy is the calculated the same way as policy iteration with the estimated optimal value function. According to the book by Sutton and Barto, value iteration can be comprehended this way: the loop of value iteration is truncated with only one iteration, and also policy improvement is done only once at the end.

As I repeated, I think policy iteration is more than just a single algorithm. And relations of values and policies should be discussed carefully rather than just following pseudocode. And whatever RL algorithms you learn, I think more or less you find some similarities to policy iteration. Thus in the next article, I would like to introduce policy iteration in more abstract ways. And I am going to take a rough look at various major RL algorithms with the keywords of “values” and “policies” in the next article.


I mentioned the Bellman equation is nothing obvious. In this section, I am going to introduce a mathematical derivation, which I think is the most straightforward. If you are allergic to mathematics, the part blow is not recommendable, but the Bellman equation is the core of RL. I would not say this is difficult, and if you are going to read some texts on RL including some equations, I think mastering the operations I explain below is almost mandatory.

First of all, let’s organize some important points. But please tolerate inaccuracy of mathematical notations here. I am going to follow notations in the book by Sutton and Barto.

  • Capital letters usually denote random variables. For example X, Y,Z, S_t, A_t, R_{t+1}, S_{t+1}. And corresponding small letters are realized values of the random variables. For example x, y, z, s, a, r, s'. (*Please do not think too much about the number of 's on the small letters.)
  • Conditional probabilities in general are denoted as for example \text{Pr}\{X=x, Y=y | Z=z\}. This means the probability of x, y are sampled given that z is sampled.
  • In the book by Sutton and Barto, a probilistic funciton p(\cdot) means a probability of transition, but I am using p(\cdot) to denote probabilities in general. Thus p( s', a, r | s) shows the probability that, given an agent being in state s at time t, the agent will do action a, AND doing this action will cause the agent to proceed to state s' at time t+1, and receive reward r. p( s', a, r | s) is not defined in the book by Barto and Sutton.
  • The following equation holds about any conditional probabilities: p(x, y|z) = p(x|y, z)p(y|z). Thus importantly, p(s', a, r|s) = p(s', r| s, a)p(a|s)=p(s', r | s, a)\pi(a|s)
  • When random variables X, Y are discrete random variables, a conditional expectation of X given Y=y is calculated as follows: \mathbb{E}[X|Y=y] = \sum_{x}{p(x|Y=y)}.

Keeping the points above in mind, let’s get down on business. First, according to definition of a value function on a policy pi and linearity of an expectation, the following equations hold.

v_{\pi}(s) = \mathbb{E} [G_t | S_t =s] = \mathbb{E} [R_{t+1} + \gamma G_{t+1} | S_t =s]

=\mathbb{E} [R_{t+1} | S_t =s] + \gamma \mathbb{E} [G_{t+1} | S_t =s]

Thus we need to calculate \mathbb{E} [R_{t+1} | S_t =s] and \mathbb{E} [G_{t+1} | S_t =s]. As I have explained \mathbb{E} [R_{t+1} | S_t =s] is the sum of p(s', a, r |s) r over all the combinations of (s', a, r). And according to one of the points above, p(s', a, r |s) = p(s', r | s, a)p(a|s)=p(s', r | s, a)\pi(a|s). Thus the following equation holds.

\mathbb{E} [R_{t+1} | S_t =s] = \sum_{s', a, r}{p(s', a, r|s)r} = \sum_{s', a, r}{p(s', r | s, a)\pi(a|s)r}.

Next we have to calculate

\mathbb{E} [G_{t+1} | S_t =s]

= \mathbb{E} [R_{t + 2} + \gamma R_{t + 3} + \gamma ^2 R_{t + 4} + \cdots | S_t =s]

= \mathbb{E} [R_{t + 2}  | S_t =s] + \gamma \mathbb{E} [R_{t + 2} | S_t =s]  + \gamma ^2\mathbb{E} [ R_{t + 4} | S_t =s]  +\cdots.

Let’s first calculate \mathbb{E} [R_{t + 2}  | S_t =s]. Also \mathbb{E} [R_{t + 3}  | S_t =s] is a sum of p(s'', a', r', s', a, r|s)r' over all the combinations of (s”, a’, r’, s’, a, r).

\mathbb{E}_{\pi} [R_{t + 2}  | S_t =s] =\sum_{s'', a', r', s', a, r}{p(s'', a', r', s', a, r|s)r'}

=\sum_{s'', a', r', s', a, r}{p(s'', a', r'| s', a, r, s)p(s', a, r|s)r'}

=\sum_{ s', a, r}{p(s', a, r|s)} \sum_{s'', a', r'}{p(s'', a', r'| s', a, r, s)r'}

I would like you to remember that in Markov decision process the next state S_{t+1} and the reward R_t only depends on the current state S_t and the action A_t at the time step.

Thus in variables s', a, r, s, only s' have the following variables r', a', s'', r'', a'', s''', \dots.  And again p(s', a, r |s) = p(s', r | s, a)p(a|s). Thus the following equations hold.

\mathbb{E}_{\pi} [R_{t + 2}  | S_t =s]=\sum_{ s', a, r}{p(s', a, r|s)} \sum_{s'', a', r'}{p(s'', a', r'| s', a, r', s)r'}

=\sum_{ s', a, r}{p(s', r|a, s)\pi(a|s)} \sum_{s'', a', r'}{p(s'', a', r'| s')r'}

= \sum_{ s', a, r}{p(s', r|a, s)\pi(a|s)} \mathbb{E}_{\pi} [R_{t+2}  | s'].

\mathbb{E}_{\pi} [R_{t + 3}  | S_t =s] can be calculated the same way.

\mathbb{E}_{\pi}[R_{t + 3}  | S_t =s] =\sum_{s''', a'', r'', s'', a', r', s', a, r}{p(s''', a'', r'', s'', a', r', s', a, r|s)r''}

=\sum_{s''', a'', r'', s'', a', r', s', a, r}{p(s''', a'', r'', s'', a', r'| s', a, r, s)p(s', a, r|s)r''}

=\sum_{ s', a, r}{p(s', a, r|s)} \sum_{s''', a'' r'', s'', a', r'}{p(s''', a'', r'', s'', a', r'| s', a, r, s)r''}

=\sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} \sum_{s''', a'' r'', s'', a', r'}{p(s''', a'', r'', s'', a', r'| s')r''}

=\sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} \mathbb{E}_{\pi} [R_{t+3}  | s'].

The same is true of calculating \mathbb{E}_{\pi} [R_{t + 4}  | S_t =s], \mathbb{E}_{\pi} [R_{t + 5}  | S_t =s]\dots.  Thus

v_{\pi}(s) =\mathbb{E} [R_{t+1} | S_t =s] + \gamma \mathbb{E} [G_{t+1} | S_t =s]

=\sum_{s', a, r}{p(s', r | s, a)\pi(a|s)r} + \mathbb{E} [R_{t + 2}  | S_t =s] + \gamma \mathbb{E} [R_{t + 3} | S_t =s]  + \gamma ^2\mathbb{E} [ R_{t + 4} | S_t =s]  +\cdots

=\sum_{s, a, r}{p(s', r | s, a)\pi(a|s)r} +\sum_{ s', a, r}{p(s', r|a, s)\pi(a|s)} \mathbb{E}_{\pi} [R_{t+2}  |S_{t+1}= s'] +\gamma \sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} \mathbb{E}_{\pi} [R_{t+3} |S_{t+1} =  s'] +\gamma^2 \sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} \mathbb{E}_{\pi} [ R_{t+4}|S_{t+1} =  s'] + \cdots

=\sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} [r + \mathbb{E}_{\pi} [\gamma R_{t+2}+ \gamma R_{t+3}+\gamma^2R_{t+4} + \cdots |S_{t+1} =  s'] ]

=\sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} [r + \mathbb{E}_{\pi} [G_{t+1} |S_{t+1} =  s'] ]

=\sum_{ s', a, r}{ p(s', r | s, a)p(a|s)} [r + v_{\pi}(s') ]