CSE510 Deep Reinforcement Learning (Lecture 11)

Materials Used

Much of the material and slides for this lecture were taken from Chapter 13 of Barto & Sutton textbook.

Some slides are borrowed from Rich Sutton’s RL class and David Silver’s Deep RL tutorial

Problem: often the feature-based policies that work well (win games, maximize utilities) aren’t the ones that approximate V / Q best

Q-learning’s priority: get Q-values close (modeling)
Action selection priority: get ordering of Q-values right (prediction)

Value functions can often be much more complex to represent than the corresponding policy

Do we really care about knowing $Q(s, \text{left}) = 0.3554, Q(s, \text{right}) = 0.533$ Or just that “right is better than left in state $s$ ”

Motivates searching directly in a parameterized policy space

Bypass learning value function and “directly” optimize the value of a policy

Examples

Rock-Paper-Scissors

Two-player game of rock-paper-scissors
- Scissors beats paper
- Rock beats scissors
- Paper beats rock
Consider policies for iterated rock-paper-scissors
- A deterministic policy is easily exploited
- A uniform random policy is optimal (i.e., Nash equilibrium)

Partial Observable GridWorld

Partial Observable Grid World

The agent cannot differentiate the grey state

Consider features of the following form (for all $N,E,S,W$ actions):

\phi(s,a)=1(\text{wall to} N, a=\text{move} E)

Compare value-based RL, suing an approximate value function

Q_\theta(s,a) = f(\phi(s,a),\theta)

To policy-based RL, using a parameterized policy

\pi_\theta(s,a) = g(\phi(s,a),\theta)

Under aliasing, an optimal deterministic policy will either

move $W$ in both grey states (shown by red arrows)
move $E$ in both grey states

Either way, it can get stuck and never reach the money

Value-based RL learns a near-deterministic policy
- e.g. greedy or $\epsilon$ -greedy

So it will traverse the corridor for a long time

An optimal stochastic policy will randomly move $E$ or $W$ in grey cells.

\pi_\theta(\text{wall to }N\text{ and }S, \text{move }E) = 0.5\\ \pi_\theta(\text{wall to }N\text{ and }S, \text{move }W) = 0.5

It will reach the goal state in a few steps with high probability

Policy-based RL can learn the optimal stochastic policy

RL via Policy Gradient Ascent

The policy gradient approach has the following schema:

Select a space of parameterized policies (i.e., function class)
Compute the gradient of the value of current policy wrt parameters
Move parameters in the direction of the gradient
Repeat these steps until we reach a local maxima

So we must answer the following questions:

How should we represent and evaluate parameterized policies?
How can we compute the gradient?

Policy learning objective

Goal: given policy $\pi_\theta(s,a)$ with parameter $\theta$ , find best $\theta$

In episodic environments we can use the start value:

J_1(\theta) = V^{\pi_\theta}(s_1)=\mathbb{E}_{\pi_\theta}[v_1]

In continuing environments we can use the average value:

J_{avV}(\theta) = \sum_{s\in S} d^{\pi_\theta}(s) V^{\pi_\theta}(s)

Or the average reward per time-step

J_{avR}(\theta) = \sum_{s\in S} d^{\pi_\theta}(s) \sum_{a\in A} \pi_\theta(s,a) \mathcal{R}(s,a)

here $d^{\pi_\theta}(s)$ is the stationary distribution of Markov Chain for policy $\pi_\theta$ .

Policy optimization

Policy based reinforcement learning is an optimization problem

Find $\theta$ that maximises $J(\theta)$

Some approaches do not use gradient

Hill climbing
Simplex / amoeba / Nelder Mead
Genetic algorithms

Greater efficiency often possible using gradient

Gradient descent
Conjugate gradient
Quasi-newton

We focus on gradient descent, many extensions possible

And on methods that exploit sequential structure

Policy gradient

Let $J(\theta)$ be any policy objective function

Policy gradient algorithms search for a local maxima in $J(\theta)$ by ascending the gradient of the policy with respect to $\theta$

\Delta \theta = \alpha \nabla_\theta J(\theta)

Where $\nabla_\theta J(\theta)$ is the policy gradient.

\nabla_\theta J(\theta) = \begin{pmatrix} \frac{\partial J(\theta)}{\partial \theta_1} \\ \frac{\partial J(\theta)}{\partial \theta_2} \\ \vdots \\ \frac{\partial J(\theta)}{\partial \theta_n} \end{pmatrix}

and $\alpha$ is the step-size parameter.

Policy gradient methods

The main method we will introduce is Monte-Carlo policy gradient in Reinforcement Learning.

Score Function

Assume the policy $\pi_\theta$ is differentiable and non-zero and we know the gradient $\nabla_\theta \pi_\theta(s,a)$ for all $s\in S$ and $a\in A$ .

We can compute the policy gradient analytically

We define the Likelihood ratio as:

\begin{aligned} \nabla_\theta \pi_\theta(s,a) = \pi_\theta(s,a) \frac{\nabla_\theta \pi_\theta(s,a)}{\pi_\theta(s,a)} \\ &= \nabla_\theta \log \pi_\theta(s,a) \end{aligned}

The Score Function is:

\nabla_\theta \log \pi_\theta(s,a)

Example

Take the softmax policy as example:

Weight actions using the linear combination of features $\phi(s,a)^\top\theta$ :

Probability of action is proportional to the exponentiated weights:

\pi_\theta(s,a) \propto \exp(\phi(s,a)^\top\theta)

The score function is

\begin{aligned} \nabla_\theta \ln\left[\frac{\exp(\phi(s,a)^\top\theta)}{\sum_{a'\in A}\exp(\phi(s,a')^\top\theta)}\right] &= \nabla_\theta(\ln \exp(\phi(s,a)^\top\theta) - (\ln \sum_{a'\in A}\exp(\phi(s,a')^\top\theta))) \\ &= \nabla_\theta\left(\phi(s,a)^\top\theta -\frac{\phi(s,a)\sum_{a'\in A}\exp(\phi(s,a')^\top\theta)}{\sum_{a'\in A}\exp(\phi(s,a')^\top\theta)}\right) \\ &=\phi(s,a) - \sum_{a'\in A} \prod_\theta(s,a') \phi(s,a') &= \phi(s,a) - \mathbb{E}_{a'\sim \pi_\theta(s,a')}[\phi(s,a')] \end{aligned}

In continuous action spaces, a Gaussian policy is natural

Mean is a linear combination of state features $\mu(s) = \phi(s)^\top\theta$

Variance may be fixed $\sigma^2$ , or can also parametrized

Policy is Gaussian, $a \sim N (\mu(s), \sigma^2)$

The score function is

\nabla_\theta \log \pi_\theta(s,a) = \frac{(a - \mu(s)) \phi(s)}{\sigma^2}

Policy Gradient Theorem

For any differentiable policy $\pi_\theta(s,a)$ ,

for any of the policy objective function $J=J_1, J_{avR},$ or $\frac{1}{1-\gamma}J_{avV}$ , the policy gradient is:

\nabla_\theta J(\theta) = \mathbb{\pi_\theta}[\nabla_\theta \log \pi_\theta(s,a) Q^{\pi_\theta}(s,a)]

Proof

Take $\phi(s)=\sum_{a\in A} \nabla_\theta \pi_\theta(a|s)Q^{\pi}(s,a)$ to simplify the proof.

\begin{aligned} \nabla_\theta V^{\pi}(s)&=\nabla_\theta \left(\sum_{a\in A} \pi_\theta(a|s)Q^{\pi}(s,a)\right) \\ &=\sum_{a\in A} \left(\nabla_\theta \pi_\theta(a|s)Q^{\pi}(s,a) + \pi_\theta(a|s) \nabla_\theta Q^{\pi}(s,a)\right) \text{by linear approximation}\\ &=\sum_{a\in A} \left(\nabla_\theta \pi_\theta(a|s)Q^{\pi}(s,a) + \pi_\theta(a|s) \nabla_\theta \sum_{s',r\in S\times R} P(s',r|s,a) \left(r+V^{\pi}(s')\right)\right)\text{rewrite the Q-function as sum of expected rewards from actions} \\ &=\sum_{a\in A} \left(\nabla_\theta \pi_\theta(a|s)Q^{\pi}(s,a) + \pi_\theta(a|s) \sum_{s',r\in S\times R} P(s',r|s,a) \nabla_\theta \left(r+V^{\pi}(s')\right)\right) \\ &=\phi(s)+\sum_{a\in A} \left(\pi_\theta(a|s) \sum_{s'\in S} P(s'|s,a) \nabla_\theta V^{\pi}(s')\right) \\ &=\phi(s)+\sum_{s\in S} \sum_{a\in A} \pi_\theta(a|s) P(s'|s,a) \nabla_\theta V^{\pi}(s') \\ &=\phi(s)+\sum_{s\in S} \rho(s\to s',1)\nabla_\theta V^{\pi}(s') \text{ notice the recurrence relation}\\ &=\phi(s)+\sum_{s'\in S} \rho(s\to s',1)\left[\phi(s')+\sum_{s''\in S} \rho(s'\to s'',1)\nabla_\theta V^{\pi}(s'')\right] \\ &=\phi(s)+\left[\sum_{s'\in S} \rho(s\to s',1)\phi(s')\right]+\left[\sum_{s''\in S} \rho(s\to s'',2)\nabla_\theta V^{\pi}(s'')\right] \\ &=\cdots\\ &=\sum_{x\in S}\sum_{k=0}^\infty \rho(s\to x,k)\phi(x) \end{aligned}

Just to note that $\rho(s\to x,k)=\sum_{a\in A} \pi_\theta(a|s) P(x|s,a)^k$ is the probability of reaching state $x$ in $k$ steps from state $s$ .

Let $\eta(s)=\sum_{k=0}^\infty \rho(s_0\to s,k)$ be the expected number of steps to reach state $s$ from state $s_0$ .

Note that $\sum_{s\in S} \eta(s)$ is constant depends solely on the initial state $s_0$ and policy $\pi_\theta$ .

So $d^{\pi_\theta}(s)=\frac{\eta(s)}{\sum_{s'\in S} \eta(s')}$ is the stationary distribution of the Markov Chain for policy $\pi_\theta$ .

\begin{aligned} \nabla_\theta J(\theta)&=\nabla_\theta V^{\pi}(s_0)\\ &=\sum_{s\in S} \sum_{k=0}^\infty \rho(s_0\to s,k)\phi(s)\\ &=\sum_{s\in S} \eta(s)\phi(s)\\ &=\sum_{s\in S} \eta(s)\sum_{a\in A} \frac{\eta(s)}{\sum_{s'\in S} \eta(s')}\phi(s)\\ &\propto \sum_{s\in S} \frac{\eta(s)}{\sum_{s'\in S} \eta(s')}\phi(s)\\ &=\sum_{s\in S} d^{\pi_\theta}(s)\sum_{a\in A} \nabla_\theta \pi_\theta(a|s)Q^{\pi_\theta}(s,a)\\ &=\left[\sum_{s\in S} d^{\pi_\theta}(s)\sum_{a\in A} \pi_\theta(a|s)\right]\nabla_\theta Q^{\pi_\theta}(s,a)\\ &= \mathbb{E}_{\pi_\theta}[\nabla_\theta \log \pi_\theta(s,a) Q^{\pi_\theta}(s,a)] \end{aligned}