This is a 3 part series of Deep Q-Learning, which is written such that undergrads with highschool maths should be able to understand and hit the ground running on their deep learning projects. This series is really just the literature review section of my final year report (which in on Deep Q-Learning) broken to 3 chunks:

- Part 1 – Convolutional Neural Networks
- Part 2 – Reinforcement Learning
- Part 3 – Deep Q-Learning

## Prologue: A Brief History

In as early as 1954, scientists began to speculate that psychological principles in learning will become important to building artificial intelligent systems (Minsky, 1954). The science and study of teaching computers to learn from experience become eventually known as reinforcement learning (RL).

Successful implementations of RL algorithms began in 1959 when Arthur Samuel published his work on an AI that can master checkers simply by playing with itself (Samuel, 1959). In this seminal work, he pioneered various ideas on RL, one of which spawn to a class of learning techniques known as temporal difference (TD) learning. Later in 1992, Gerry Tesauro had similar success in the game backgammon with his program, TD-Gammon (Tesauro, 1995), which surpassed all previous computer programs in its ability to play backgammon. What is new in Tesauro’s work is that TD-Gammon combines TD learning with another AI technique that is inspired by biological brains: a neural network.

However, in spite of these advances, in both Samuel’s checker program and TD-Gammon, the input is distilled to what the computer can understand (board positions), and the rules are hard-coded to the machine. This no longer becomes the case when DeepMind publishes a novel technique known as Deep Q-Learning. It made its official debut by playing retro console games off an Atari 2600 emulator (Mnih et al., 2013), and in certain games the AI is even able to outmatch expert human players. What is significant about Deep Q-Learning is that the AI learns by seeing the same output on the game console as human players do, and figures out the rules of the game without supervision. As with TD-Gammon, Deep Q-Learning also used an artificial neural network, albeit a different architecture – convolutional neural network (CNN).

## Introduction

In all the hype about machine learning in this era, some people call reinforcement learning (RL) a hybrid between supervised and unsupervised learning. Some people place reinforcement learning in a different field altogether, because knowing supervised and unsupervised learning does not mean one would understand reinforcement learning, and vice versa.

The core idea of RL comes natural to us even as an infant. We are born in a world which we knew very little of, in a body which we do not choose, and with little supervision. We learn remedial tasks, such as walking and talking, by continuously trying until we get it right. An infant does not have an understanding of failure and success; he makes attempts, sees that some things work, and some don’t, and focus more of what works. He does not get depressed or frustrated even if he keeps falling when he tries to walk. Really the only people getting emotional of his attempts are his parents. As we get older we derive more unnecessarily complex views of the concept of what should work and what is success, but that is besides the point here.

RL is the study of a *computational* approach to learning by trying. We teach a computer to learn as we do: by repeatedly interacting with the environment, and learning from prior experience. For example, a robot learns to find a path out of a maze by trying out different routes that it can see. To put it in more formal terms, we have an **agent **(robot) that starts in an initial **state** (robot’s position and what it sees), take an **action** (movement in some direction) in an **environment** (the maze), which would then return a **reward** (some score to let the robot know how close it is to finishing the maze) and a new state.

Should we describe supervised learning in terms of this agent-environment interface, we could say that an action (input) comes together with the reward (target); we train a model with both the input and the expected answer. RL introduces the concept of *delayed rewards*, where the agent is only aware of the reward *after* the action is taken. This meant that during training, the agent attempts an action, but will not know if it is the best course of action (determined by a reward value) until after it has taken an action and after it has landed on a new state.

## Markov Decision Process

Markov Decision Process (MDP) introduces a mathematical framework for describing decision making in a probabilistic environment. All RL algorithms in this series will be described using MDP.

Time in the real world is continuous; but in MDP, time is discrete, described in **time steps** in fixed intervals: , where increases with each action we take. In a given time step , an MDP is defined by:

**A set of states**, where is the set of all possible states.**A set of all actions**, where is the set of all possible actions that can be taken in the state .**A transition function**which returns the probability that if at time we take an action in the state we would end up in new state in a next time step .**A reward function**, which is the consequence of the action and comes 1 time step later. This value is normally an integer, and can also be negative value (punishment).

Transition and reward functions are also written as and respectively; this is useful notation in cases where time is not a consideration. Reward also has a more succinct notation: .

In MDP, possible actions in a current state depends only on the current state. What happens in the past and what will happen in the future does not affect what actions an agent can take now.

MDPs are used to model non-deterministic (or stochastic) search problems. This is different from deterministic search because unlike deterministic problems, non-deterministic problems do not have a fixed sequence of actions from start to finish as a final answer. The fundamental problem of an MDP is to find a **policy** (this is not the same as the constant pi 3.142), where ; a policy returns an action given a state. The best solution is a policy that maximizes the sum of all rewards:

(gamma) here is called the **discount rate**, and is a float between 0 to 1 set by us to determine the present value of future rewards. The reward of the current state remains unchanged regardless of the discount rate (because anything to the power of 0 is 1), whereas future rewards will decay exponentially with each time step. If approaches 1, the agent becomes very farsighted and takes future rewards seriously; if approaches 0, the agent has tunnel vision and only values immediate rewards. In the DeepMind’s Deep Q-Learning implementation, is set to 0.99. Discount rates are important, for without it our learning algorithm will not be able to converge (or finish running).

## Utility

We now know that the goal of the agent is to decide on what actions to maximize the sum of rewards. But to decide on an action the agent needs to have an expectation of what reward it will get. This expected reward is known as the **utility** (or value). There are 2 ways we can compute a value (Marsland, 2015):

**State-value function**, – We consider the current state, and average across all of the actions that can be taken.**Action-value function**, – We consider the current state and each possible action that can be taken separately. This is also known as a Q-value. A Q-state is when you were in a state and took an action.

In either case we are thinking about what the expected reward would be if we started in state s (where is the statistical expectation):

The utility starting in a state and acting optimally is defined as (optimal state-value function); the utility starting out having taken action from state and thereafter acting optimally as defined as (optimal state-action function) (Klein, 2014). If an agent knows all these or values, it will be able to navigate through the state space such that it attains the maximum reward; in other words, our agent will always be acting optimally.

This is illustrated in the grid world example in figure above. Here the agent navigates from one box (each box represents a state) to another until it reaches a terminal state: a box with double outlines. In each state the agent chooses between 4 different directions, each of which is a possible action. Notice that the highest utility in each Q-state box is the utility of a V-state box.

## The Bellman Equations

How shall we characterize optimal utilities? We will use bellman equations. The intuition behind it is to take the correct first action, and from then onwards continue to be optimal. This is done in a recursive fashion:

In these equations, we find all possible consequent states that can result from from taking action from state . The utility of each of these consequent states () is the reward of the current state multiplied by future expected rewards of the consequent state; this is recursively denoted as . To prioritize the reward of the current state we apply a discount rate to future expected utilities. Because the utility of all consequent states are stochastic, we multiply the culminated utilities of each consequent state by the probability they would be executed ().

## Value Iteration Algorithm

Now that we know what and are, the next step is to compute the these optimal utilities. To do this we use the value iteration algorithm.

We assume there will always exist an optimal Q-value for each Q-state. At first, all Q-states will be initialized to small random values. Then we will iterate through every possible state, evaluating its future states in the process. With a proper discount rate , an optimal Q-value will eventually converge in each Q-state.

The value iteration algorithm for finding and are given in Algorithm 1 and Algorithm 2 respectively.

Notice that in the value iteration algorithm, the probability of entering a state is known. This is not usually the case in most real world environments. Also when the state space is large (or infinite), value iteration will never finish running. To circumvent these limitations, we introduce Q-Learning.

## Q-Learning

Q-Learning (Watkins, 1989) belongs to a class of RL methods called **temporal difference learning**. It offers an online algorithm to iteratively *approximate* Q-values without knowing the transitions between states.

This is done with an update function that finds the temporal difference between learning experiences. Temporal difference simply means the difference between the reward at this current step and the estimated reward we got at the previous step. The Q-values are then determined by a running average. Naturally, the more we iterate, the closer we converge to optimal values.

To do this we break down how we find our Q-values: instead of recursively computing an infinite amount of time steps, we use a one-step value iteration update given by the Bellman equation, which is just depth-one expansion of the recursive definition of a Q-value:

This will be our estimated utility for the current state. Let us call this estimated utility and simplify to . The estimated utility then becomes:

The Q-Learning algorithm is given in Algorithm 3:

In the update function, is also called **TDError** or temporal difference error. can be describe as the learning rate, as it determines the weightage between current estimates and previous estimates. To show this more clearly, notice that these update functions are identical:

As time passes, it is usually the case that we incrementally reduce , to place priority to past estimates.

Note that in Q-learning we do not use the policy to find the value of , but instead choose the one that gives the highest value. This is known as an **off-policy decision**.

## Afterword

If you want to learn more about RL, I would recommend looking up the AI course (CS188) from the university of Berkeley. I didn’t actually finish the course – just the lectures that involve RL, and attempted the corresponding programming assignments. The grid world example is taken from there; it is actually a programming assignment (python3) where you have to write the code to compute the Q values. Marsland’s Machine Learning textbook only takes up a small part to talk about RL, but the material is relatively easier to understand, not to mention that it comes with python source code. The definitive guide to RL would definitely be Sutton’s RL book, but I had a lot of difficulty understanding it, since it just teach concepts.

I understand if the material here is pretty esoteric; I took a long time to understand it myself. However, if you manage to make it thus far, you are ready to understand the Deep Q-Learning algorithm, which I will cover in the final part of this series.

## References

- Sutton, R. S., & Barto, A. G. (1998). Reinforcement learning: An introduction (Vol. 1) (No. 1). MIT press Cambridge.
- Minsky, M. L. (1954). Theory of neural-analog reinforcement systems and its application to the brain model problem. Princeton University.
- Samuel, A. L. (1959). Some studies in machine learning using the game of checkers. IBM Journal of research and development, 3(3), 210–229.
- Tesauro, G. (1995). Temporal diﬀerence learning and td-gammon. Communications of the ACM , 38(3), 58–68.
- Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013). Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602.
- Marsland, S. (2015). Machine learning: an algorithmic perspective. CRC press.
- Klein, D. (2014). Cs 188: Artiﬁcial intelligence. University of California, Berkeley.
- Watkins, C. J. C. H. (1989). Learning from delayed rewards (Unpublished doctoral dissertation). University of Cambridge England.