Top | Notations | Bibliography

Introduction

In the REINFORCE algorithm, a policy \(\pi\) is lerned that maximizes the agent objective \(J_\pi\). Some prerequisites from Introduction to Machine Learning:

A trajectory \(\tau\) denotes a sequence of states and actions \(\begin{equation} \label{eq:tau} \tau = s_0, a_0, s_1, a_1, ... \end{equation}\)

The return of a trajectory \(\tau\) that starts at step \(t\) is denoted:

\(\begin{equation} \label{eq:trajret} r_t(\tau) = r(s_{t}, a_{t}) + {\gamma}r(s_{t+1}, a_{t+1}) + ... = \sum_{t'=t}^{\infty} \gamma^{t'-t}r(s_{t'},a_{t'}) \end{equation}\)

The agent objective

The agent needs to learn a policy that maximizes the agent objective \(J_\pi\):

\[\begin{align} J_\pi & = \underset{T \rightarrow \infty}{lim} \mathbb{E}_{\tau \sim \pi}[\sum_{t=0}^{T-1} \gamma^{t} r(s_t, a_t)] & \\ & = \sum_{t=0}^{\infty} \int_{\tau_{\le a_t} = s_0, a_0, ... , a_t} \gamma^{t} r(s_t, a_t) p_\pi(s_t, a_t \vert \tau) d\tau_{\le a_t} \hspace{1cm} & \\ & = \sum_{t=0}^{\infty} \gamma^{t} \mathbb{E}_{\tau_{\le a_t} \sim \pi}[r(s_t, a_t)] & (definition \, of \, expectation) \end{align}\]

We assume that the number of steps \(T\), the state space \(\mathcal{S}\) and action space \(\mathcal{A}\) are finite. (See here how the same can be done for a continuous action space \(\mathcal{A}\).)

The deep neural network

We construct a deep neural network with states \(s \in \mathcal{S}\) as input, and policy distributions \(\pi(a \vert s)\) outputs for all actions \(a \in \mathcal{A}\).

The weights of the NN are denoted \(\theta\), and parametrize the output policy \(\pi_\theta\). The NN metric is the agent objective \(J_{\pi_\theta}\). The algorithm starts with a random policy \(\pi_\theta\), samples states \(s \in \mathcal{S}\), and uses gradient ascent to maximize \(J_{\pi_\theta}\). At each step, the NN weights are updated according to the rule \(\begin{align} \theta = \theta + \alpha \nabla_\theta J_{\pi_\theta} \end{align}\) where \(\alpha\) is the learning rate. The key, here, is to be able to compute the gradient \(\nabla_\theta J_{\pi_\theta}\).

Computing the policy gradient

Here, \(J_{\pi_\theta} = \mathbb{E}_{\tau \sim \pi_\theta}[R(\tau)]\), so, more generally, given a function \(f(x)\), and a conditional distribution \(p(x \vert \theta)\), it would help to compute the gradient of its expectation \(\mathbb{E}_{x \sim p(x \vert \theta)}[f(x)]\). We have:

\[\begin{align} \nabla_\theta \mathbb{E}_{x \sim p(x \vert \theta)}[f(x)] & = \nabla_\theta \int_x f(x) p(x \vert \theta) dx & (definition \, of \, expectation) \\ & = \int_x f(x) \nabla_\theta (p(x \vert \theta)) dx & (gradient \, commutes \, with \, the \, integral) \\ & = \int_x f(x) p(x \vert \theta) \frac{\nabla_\theta (p(x \vert \theta))}{p(x \vert \theta)} dx & (multiply \, and \, divide \, with \, p(x \vert \theta)) \\ & = \int_x f(x) p(x \vert \theta) \nabla_\theta (ln \, p(x \vert \theta)) dx & (chain \, rule) \\ & = \mathbb{E}_{x \sim p(x \vert \theta)}[f(x)\nabla_\theta (ln \, p(x \vert \theta))] & (definition \, of \, expectation) \\ \end{align}\]

In view of this,

\[\begin{align} \nabla_\theta J_{\pi_\theta} & = \nabla_\theta \sum_{t=0}^{\infty} \gamma^{t} \mathbb{E}_{\tau_{\le a_t} \sim \pi_\theta}[r(s_t, a_t)] & (express \, J_{\pi_\theta} \, as \, sum \, of \, expectations) \\ & = \sum_{t=0}^{\infty} \gamma^{t} \nabla_\theta \mathbb{E}_{\tau_{\le a_t} \sim \pi_\theta}[r(s_t, a_t)] & (bring \, \nabla_\theta \, inside \, sum) \\ & = \sum_{t=0}^{\infty} \gamma^{t} \mathbb{E}_{\tau_{\le a_t} \sim \pi_\theta}[r(s_t, a_t) \nabla_\theta (ln \, p_{\pi_\theta}(\tau_{\le a_t} \vert \theta))] & (bring \, \nabla_\theta \, inside \, \mathbb{E}_{\tau_{\le a_t} \sim \pi_\theta})\\ \end{align}\]

Recall that \(p_\pi(\tau_{\le a_t}) = d(s_0) \prod_{t'=0}^{t-1} \pi(a_{t'} \vert s_{t'}) \prod_{t'=0}^{t-1} p(s_{t'+1} \vert s_{t'}, a_{t'})\). We can expand the term \(\nabla_\theta ln \, p_\pi(\tau_{\le a_t} \vert \theta)\) as follows:

\[\begin{align} \nabla_\theta ln \, p_{\pi_\theta}(\tau_{\le a_t} \vert \theta) &= \nabla_\theta ln \, \{d(s_0) \prod_{t'=0}^t \pi_\theta(a_{t'} \vert s_{t'}) \prod_{t'=0}^t p(s_{t'+1} \vert s_{t'}, a_{t'}) \vert \theta)\} & (expand \, p_{\pi_\theta}(\tau_{\le a_t}) \\ & = \nabla_\theta ln \, d(s_0) + \sum_{t'=0}^t \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'}) + \sum_{t'=0}^t \nabla_\theta ln \, p(s_{t'+1} \vert s_{t'}, a_{t'}) & (log \, of \, product \, is \, sum \, of \, logs)) \\ & = \sum_{t'=0}^t \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'}) & (\nabla_\theta \, of \, terms \, that \, do \, not \, depend \, on \, \theta \, is \, 0)) \\ \end{align}\]

Applying that:

\[\begin{align} \nabla_\theta J_{\pi_\theta} & = \underset{T \rightarrow \infty}{lim} \sum_{t=0}^{T-1} \gamma^{t} \mathbb{E}_{\tau_{\le a_t} \sim \pi_\theta}[r(s_t, a_t) \sum_{t'=0}^t \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'})] & \\ & = \underset{T \rightarrow \infty}{lim} \sum_{t=0}^{T-1} \gamma^{t} \mathbb{E}_{\tau_{\le a_T} \sim \pi_\theta}[r(s_t, a_t) \sum_{t'=0}^t \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'})] & (expectation \, remains \, same \, when \, using \, \tau_{\le a_T} \, instead \, of \, \tau_{\le a_t})\\ & = \underset{T \rightarrow \infty}{lim} \mathbb{E}_{\tau_{\le a_T} \sim \pi_\theta}[\sum_{t=0}^{T-1} \{ \gamma^{t} r(s_t, a_t) \sum_{t'=0}^t \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'}) \}] & (bring \, outer \, sum \, in)\\ & = \underset{T \rightarrow \infty}{lim} \mathbb{E}_{\tau_{\le a_T} \sim \pi_\theta}[\sum_{0 \le t' \le t \le T} \{\gamma^{t} r(s_t, a_t) \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'}) \}] & (convert \, from \, double \, sum) \\ & = \underset{T \rightarrow \infty}{lim} \mathbb{E}_{\tau_{\le a_T} \sim \pi_\theta}[\sum_{t'=0}^{T-1} \{\nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'}) \sum_{t=t'}^{T-1} \gamma^{t} r(s_t, a_t) \}] & (convert \, to \, reverse \, double \, sum) \\ & = \underset{T \rightarrow \infty}{lim} \mathbb{E}_{\tau_{\le a_T} \sim \pi_\theta}[\sum_{t'=0}^{T-1} \{\gamma^{t'} r_{t'}(\tau) \nabla_\theta ln \, \pi_\theta(a_{t'} \vert s_{t'}) \}] & (definition \, of \, r_t(\tau)) \\ & = \underset{T \rightarrow \infty}{lim} \mathbb{E}_{\tau_{\le a_T} \sim \pi_\theta}[\sum_{t=0}^{T-1} \{\gamma^t r_t(\tau) \nabla_\theta ln \, \pi_\theta(a_{t} \vert s_{t}) \}] & (rename \, t' \, as \, t) \\ \end{align}\]

The REINFORCE algorithm

\(~~~~\) 1: Initialize learning rate \(\alpha\)
\(~~~~\) 2: Initialize policy network weights \(\theta\)
\(~~~~\) 3: for \(episode = 0,..., MAX\_EPISODE\) do
\(~~~~~~\) 4: Sample a trajectory \(\tau = s_0, a_0, ..., s_T, a_T\)
\(~~~~~~\) 5: Set \(\nabla_\theta J_{\pi_{\theta}} = \sum_{t=0}^{T-1} \{\gamma^t r_t(\tau) \nabla_\theta ln \, \pi_\theta(a_{t} \vert s_{t}) \}\)
\(~~~~~~\) 6: \(\theta = \theta + \alpha \nabla_\theta J_{\pi_{\theta}}\)

The algorithm REINFORCE learns the policy directly, thus cannot be trained on previously collected samples. REINFORCE is a model-free, on-policy algorithm.

Example: CartPole

See the annotated implementation: reinforce.py. Follow instructions in README to run. Example output:

Episode 214, loss 3671.379638671875, total_reward 175.0, solved False
Episode 215, loss 4000.337158203125, total_reward 200.0, solved True
Episode 216, loss 3653.253173828125, total_reward 200.0, solved True
Episode 217, loss 3876.510498046875, total_reward 186.0, solved False
Episode 218, loss 3042.779296875, total_reward 185.0, solved False
Episode 219, loss 4446.62890625, total_reward 200.0, solved True
Episode 220, loss 4500.43359375, total_reward 200.0, solved True
Episode 221, loss 3556.630859375, total_reward 200.0, solved True

The total possible reward per episode is 200. The game is solved if the cartpole does not fall in 195 steps out of 200. Up to 300 episodes are tried.

Vanilla Actor-Critic Algorithm

See here for example. The idea of Actor-Critic is to learn both the policy \(\pi_\theta\) and the action-value \(Q_{\pi_{\theta}}(s, a)\), and to update policy weights \(\theta\) in the direction of \(\nabla_\theta Q_{\pi_{\theta}}(s, a)\) instead of \(\nabla_\theta J_{\pi_{\theta}}\).