Skip to main content

REINFORCEMENT LEARNING- PART 2

For an introduction to Reinforcement Learning, its basic terminologies, concepts and types read Reinforcement Learning - Part 1 by following this link: http://blog.cerelabs.com/2017/04/reinforcement-learning-part-1.html
 

Q-Learning

Q learning is an algorithm in reinforcement learning. It originates from the model based reinforcement learning. It can be referred to as a different kind of value function. The values are called Q values and are denoted by Q(s,a). It signifies the Q value when in a state 's' and taking an action 'a'.

Mathematically,

                      Q(s,a) = R(s) + γ Σs' P(s,a,s') maxa' Q(s',a')


It can be defined as the value for arriving in a state which is obtained by learning via action 'a' and proceeding optimally thereafter.

Also,
                           V(s)     = maxa Q(s,a)
                                 л(s)      =   argmaxa Q(s,a)

V(s) is a value, i.e. it returns a number , a scalar value in particular, whereas л(s) returns an action. Hence, it is written as argmax.

V(s) = The reward and behaving optimally thereafter.
Q(s) = The reward and then taking a specific action and behaving optimally thereafter.

We can convert Q into V if we always pick the best action.

Therefore,
                           V(s) = maxa Q(s,a)

л(s) is the policy we want to follow that maximizes your value going towards the wall, except the difference is, it is returning an action not an value, so it should be argmax.

                           л(s) = argmaxa Q(s,a)

To find Q in Q-learning

To solve the Q(s,a) equation we don't have the P(s,a,s') values. But in the learning scenario we don't have the models we have the transitions.

Estimating Q from transitions,
                 Q(s,a)  =  R(s) + γ Σs' P(s,a,s') maxa' Q(s',a')

A transition means: We observe that we are in a particular state s of the MDP, then action 'a' was chosen somehow, then transition happens and we land in a state and we find out what state we are in.

Now, imagine we got an estimate of the Q function Q^(s,a), and we are going to update it using learning rate alpha.
          Q^(s,a) <-- α --- { R(s) + γ maxa' Q^(s',a')} is utility of a state

Where,
α = learning rate
R(s) = Immediate reward
γ maxa' Q^(s',a') = discounted estimated value of the next state (utility of next state)

With each iteration Q^(s,a) approaches towards the sum of the immediate reward and discounted estimated value of the next state. This is also called as utility of state.

eg.
V <== α == X means moving alpha percent of value X to V with each iteration.
                                              
                 V = (1- α V) + αX

As we know s , a, s' ,r,
                 Q^(s,a) <--αt--- R(s) + γ maxa' Q^(s',a')

Where,
αt = learning rate increasing over time.
                 Q^(s,a) = E { R(s) + γ maxa' Q^(s',a')}
                                    = R(s) + γ Es' [ maxa' Q^(s',a') ]
                                    = R(s) + γ Σs' P(s,a,s') Q^(s',a')


The value of Q^(s,a) actually approaches to the value of the 'utility of state' term. But, the term value is changing over time, hence it is a moving target. There is a theorem by which this equation can solve the Markov Decision Process, it is called the Q- learning convergence theorem.

Q-Learning Convergence
In this theorem, Q^(s,a) starts at a random value. Then with each iteration, the value of Q^(s,a) converges to Q(s,a). The speed of convergence depends on learning rate alpha.

                    Q^(s,a) <--αt--- R(s) + γ maxa' Q^(s',a')
                    Q^(s,a) ---converges to -------> Q(s,a)

This is the actual solution to Bellman's equation. But, it is only true if we visit (s,a) infinitely often.

Choosing Actions

Q-learning is a family of algorithms. In order to solve the above problem we need to find three things:
  • How to initialize Q^?
  • How to decay α ?
  • How to choose actions?

There are two options of choosing actions:
  1. Always Choose a0
  2. Choose 'a' randomly, then take all a's and find best Q(s,a)

In the second option of choosing randomly we might learn the best action to take to get best Q(s,a) but we don't use it. So, the next step is to use it.

Now we initialize,
                Q^(s , a0 )  =  awesome;    for all states
                                  and
                    Q^(s , a  )  =  terrible ;       for all actions except a0

So, when it goes to a0 , it gets awesome results. But, whenever it goes to any other 'a' then it gets really really bad results. Hence, it continues to use a0 action forever i.e. infinitely. This is known as greedy action selection strategy.

So, we have to be careful where we start and then take random restarts. But this process could be very slow. So, instead of taking random restarts we can just take random action once in a while. Once in a state, find the estimated value for each action and take that with probability ( 1 - ε).

Where,
ε = randomness factor
                 л^(s) = argmaxa Q^(s,a) ;         for P =( 1 - ε )
                                 and
                       Random Action                            for P = ε

Assuming 'ε ' is small.
Hence, we visit (s,a) an infinite number of times as long as (ε > 0 ).


ε- Greedy Exploration:
If action selection is GLIE ( Greedy in limit with infinite exploration ). This means we start off 'ε' randomly then get less and less random and more and more greedy. We get two facts from this

                  Q^ -----> Q        <==== Learn
                  л^ -----> л          <==== Use
This is an example of Exploration- Exploitation dilemma.

Summary
This part, therefore gives a brief idea of Q-learning algorithm, q-values and their mathematical representation. Until now, the reinforcement learning robot has observed its present state and found out the q-values for all possible states in future. Using Q-value convergence theorem we find the appropriate q-values and using ε- greedy exploration take proper action. This part therefore covers up how the robot works conceptually. The next and the final part will show how the robot interacts with the real world.


References:

  
By Sanjuksha Nirgude,  
Research Project Fellow,
Cere Labs Pvt. Ltd



Comments

Popular posts from this blog

Understanding Generative Adversarial Networks - Part II

In "Understanding Generative Adversarial Networks - Part I" you gained a conceptual understanding of how GAN works. In this post let us get a mathematical understanding of GANs.
The loss functions can be designed most easily using the idea of zero-sum games. 
The sum of the costs of all players is 0. This is the Minimax algorithm for GANs
Let’s break it down.
Some terminology: V(D, G) : The value function for a minimax game E(X) : Expectation of a random variable X, also equal to its average value D(x) : The discriminator output for an input x from real data, represents probability G(z): The generator's output when its given z from the noise distribution D(G(z)): Combining the above, this represents the output of the discriminator when 
given a generated image G(z) as input
Now, as explained above, the discriminator is the maximizer and hence it tries to 
maximize

Understanding Generative Adverserial Networks - Part 1

This is a two part series on understanding Generative Adversarial Networks (GANs). This part deals with the conceptual understanding of GANs. In the second part we will try to understand the mathematics behind GANs.

Generative networks have been in use for quite a while now. And so have discriminative networks. But only in 2014 did someone get the brilliant idea of using them together. These are the generative adversarial networks. This kind of deep learning model was invented by Ian Goodfellow. When we work with data already labelled, it’s called supervised learning. It’s much easier compared to unsupervised learning, which has no predefined labels, making the task more vague. 

"Generative Adversarial Networks is the most interesting idea in the last ten years in Machine Learning." - Yann LeCun

In this post, we’ll discuss what GANs are and how they work, at a higher , more abstract level. Since 2014, many variations of the traditional GAN have come out, but the underlying conc…