Open In App

REINFORCE Algorithm

Last Updated : 05 Jun, 2025
Comments
Improve
Suggest changes
Like Article
Like
Report

REINFORCE is a method used in reinforcement learning to improve how decisions are made. It learns by trying actions and then adjusting the chances of those actions based on the total reward received afterward.

Unlike other methods that estimate how good each action is REINFORCE directly learns the best way to choose actions. This makes it especially useful for tasks where there are many possible actions or continuous choices and when it is hard to estimate the value of each action.

How REINFORCE Works

The REINFORCE algorithm works in the following steps:

  • Collect Episodes: The agent interacts with the environment for a fixed number of steps or until an episode is complete, following the current policy. This generates a trajectory consisting of states, actions and rewards.
  • Calculate Returns: For each time step t, calculate the return G_t​ which is the total reward obtained from time t onwards. Typically, this is the discounted sum of rewards:

G_t = \sum_{k=t}^T \gamma^{k-t}

Where \gamma is the discount factor, T is the final time step of the episode and R_k​ is the reward received at time step k.

  • Policy Gradient Update: The policy parameters θ are updated using the following formula:

\theta_{t+1} = \theta_t + \alpha \nabla_{\theta} \log \pi_{\theta}(a_t | s_t) G_t

Where:

α is the learning rate.

\pi_{\theta}(a_t | s_t) is the probability of taking action a_t​ at state s_t, according to the policy.

G_t is the return or cumulative reward obtained from time step t onwards.

The gradient \nabla_{\theta} \log \pi_{\theta}(a_t | s_t) represents how much the policy probability for action a_t​ at state s_t should be adjusted based on the obtained return.

  • Repeat: This process is repeated for several episodes, iteratively updating the policy in the direction of higher rewards.

REINFORCE algorithm Implementation

In this example we will train a policy network to solve a basic environment such as CartPole from OpenAI's gym. The aim is to use REINFORCE to directly optimize the policy without using value function approximations.

Step 1: Set Up the Environment

The first step is to create the environment using OpenAI's Gym. For this example we use the CartPole-v1 environment where the agent's task is to balance a pole on a cart.

Python
import gym
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers

env = gym.make('CartPole-v1')
obs_space = env.observation_space.shape[0]
act_space = env.action_space.n

Step 2: Define Hyperparameters

In this step we define hyperparameters for the algorithm like discount factor gamma, the learning rate, number of episodes and batch size. These hyperparameters control how the algorithm behaves during training.

Python
gamma = 0.99
learning_rate = 0.01
num_episodes = 1000
batch_size = 64

Step 3: Define the Policy Network (Actor)

We define the policy network as a simple neural network with two dense layers. The input to the network is the state and the output is a probability distribution over the actions (softmax output). The network learns the policy that maps states to action probabilities.

Python
class PolicyNetwork(tf.keras.Model):
    def __init__(self, hidden_units=128):
        super(PolicyNetwork, self).__init__()
        self.dense1 = layers.Dense(hidden_units, activation='relu')
        self.dense2 = layers.Dense(env.action_space.n, activation='softmax')  

    def call(self, state):
        x = self.dense1(state)
        return self.dense2(x)

Step 4: Initialize the Policy and Optimizer

Here, we initialize the policy network and the Adam optimizer. The optimizer is used to update the weights of the policy network during training.

Python
policy = PolicyNetwork()
optimizer = tf.keras.optimizers.Adam(learning_rate)

Step 5: Compute Returns

In reinforcement learning, the return G_t is the discounted sum of future rewards. This function computes the return for each time step t, based on the rewards collected during the episode.

Python
def compute_returns(rewards, gamma):
    returns = np.zeros_like(rewards, dtype=np.float32)
    running_return = 0
    for t in reversed(range(len(rewards))):
        running_return = rewards[t] + gamma * running_return
        returns[t] = running_return
    return returns

Step 6: Define Training Step

The training step computes the gradients of the policy network using the log of action probabilities and the computed returns. The loss is the negative log-likelihood of the actions taken, weighted by the return. The optimizer updates the policy network’s parameters to maximize the expected return.

Python
def train_step(states, actions, returns):
    with tf.GradientTape() as tape:
        # Calculate the probability of each action taken
        action_probs = policy(states)
        action_indices = np.array(actions, dtype=np.int32)
        
        # Gather the probabilities for the actions taken
        action_log_probs = tf.math.log(tf.reduce_sum(action_probs * tf.one_hot(action_indices, env.action_space.n), axis=1))

        # Calculate the loss (negative log likelihood * returns)
        loss = -tf.reduce_mean(action_log_probs * returns)

    grads = tape.gradient(loss, policy.trainable_variables)
    optimizer.apply_gradients(zip(grads, policy.trainable_variables))

Step 7: Training Loop

The training loop collects experiences from episodes and then performs training in batches. The policy is updated after each batch of experiences. In each episode, we record the states, actions and rewards and then compute the returns. The policy is updated based on these returns.

Python
for episode in range(num_episodes):
    state, _ = env.reset()
    done = False
    states, actions, rewards = [], [], []

    while not done:
        state_input = np.array(state, dtype=np.float32).reshape(1, -1)
        probs = policy(state_input).numpy()[0]
        action = np.random.choice(act_space, p=probs)

        next_state, reward, terminated, truncated, _ = env.step(action)
        done = terminated or truncated

        states.append(state_input[0])
        actions.append(action)
        rewards.append(reward)
        state = next_state

    # After episode ends
    returns = compute_returns(rewards, gamma)
    returns = (returns - np.mean(returns)) / (np.std(returns) + 1e-9)

    states_batch = np.vstack(states)
    train_step(states_batch, actions, returns)

    if episode % 100 == 0:
        print(f"Episode {episode}/{num_episodes}")

Step 8: Testing the Trained Agent

After training the agent, we evaluate its performance by letting it run in the environment without updating the policy. The agent chooses actions based on the highest probabilities (greedy behavior).

Python
state, _ = env.reset()
done = False
total_reward = 0

while not done:
    state_input = np.array(state, dtype=np.float32).reshape(1, -1)
    probs = policy(state_input).numpy()[0]
    action = np.argmax(probs)

    next_state, reward, terminated, truncated, _ = env.step(action)
    done = terminated or truncated
    total_reward += reward
    state = next_state

print(f"Test Total Reward: {total_reward}")

Output:

Episode 0/1000
Episode 100/1000
Episode 200/1000
Episode 300/1000
Episode 400/1000
Episode 500/1000
Episode 600/1000
Episode 700/1000
Episode 800/1000
Episode 900/1000
Test Total Reward: 49.0

Advantages of REINFORCE

  • Easy to Understand: REINFORCE is simple and easy to use and a good way to start learning about how to improve decision in reinforcement learning.
  • Directly Improves Decisions: It works by directly improving the way actions are chosen which is helpful when there are many possible actions or choices.
  • Good for Tasks with Clear Endings: It works well when tasks have a clear finish and the agent gets a total reward at the end.

Challenges of REINFORCE

  • High Variance: One of the major issues with REINFORCE is its high variance. The gradient estimate is based on a single trajectory and the return G_t​ can fluctuate significantly, making the learning process noisy and slow.
  • Sample Inefficiency: Since REINFORCE requires complete episodes to update the policy, it tends to be sample-inefficient. The agent may have to spend a lot of time trying things out before it gets helpful feedback to learn from.
  • Convergence Issues: Because the results can be very random and learning is slow REINFORCE needs a lot of practice before it learns a good way to act.

Variants of REINFORCE

Several modifications to the original REINFORCE algorithm have been proposed to address its high variance:

  • Baseline: By subtracting a baseline value (typically the value function V(s)) from the return G_t​, the variance of the gradient estimate can be reduced without affecting the expected gradient. This results in a variant known as REINFORCE with a baseline.

The update rule becomes:

\theta_{t+1} = \theta_t + \alpha \nabla_{\theta} \log \pi_{\theta}(a_t | s_t) (G_t - b_t)

Where b_t​ is the baseline such as the expected reward from state s_t​.

  • Actor-Critic: It is a method that use two parts to learn better: the actor and the critic. The actor chooses what action to take while the critic checks how good that action was and give feedback. This helps to make learning more stable and faster by reducing random mistakes.

Applications of REINFORCE

REINFORCE has been applied in several domains:

  • Robotics: REINFORCE helps robots to learn how to do things like picking up objects or moving around. The robot try different actions and learn from what works well or not.
  • Game AI: It is used to teach game players like in video games or board games like chess. The player learns by playing the game many times and figure out what moves led to win.
  • Self-driving cars: REINFORCE can help improve how self-driving cars decide to drive safely and efficiently by rewarding good driving decisions.

Next Article

Similar Reads