On-Policy Training

In online reinforcement learning, an agent is able to gather data by directly interacting with its environment. It can then use this experience to learn from and update its policy. To enable our agent to interact in this way, the agent needs to act either in the real world, or in a simulation.

AgileRL’s online training framework enables agents to learn in environments, using the standard Gym interface, 10x faster than SOTA by using our Evolutionary Hyperparameter Optimization algorithm.

On-policy reinforcement learning involves learning from experiences gathered by following a single policy. In these algorithms, the data collection policy and the learning policy are the same, meaning that the agent learns from its own actions and their outcomes in the environment. This approach often leads to more stable learning as the agent directly interacts with the environment based on its current policy, continuously updating and improving it. However, on-policy algorithms can be less sample-efficient compared to off-policy methods since they are constrained to learn from the data generated by the current policy, potentially limiting exploration and the use of past experiences.

Algorithms	Tutorials
PPO	Acrobot, Masked Velocity LunarLander-v3

Population Creation

To perform evolutionary HPO, we require a population of agents. Individuals in this population will share experiences but learn individually, allowing us to determine the efficacy of certain hyperparameters. Individual agents which learn best are more likely to survive until the next generation, and so their hyperparameters are more likely to remain present in the population. The sequence of evolution (tournament selection followed by mutation) is detailed further below.

Example Population Creation

import torch
from agilerl.utils.utils import (
    create_population,
    make_vect_envs,
    observation_space_channels_to_first
)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

NET_CONFIG = {
    "encoder_config": {"hidden_size": [32, 32]}  # Actor head hidden size
}

INIT_HP = {
    "POP_SIZE": 6,  # Population size
    "BATCH_SIZE": 128,  # Batch size
    "LR": 1e-3,  # Learning rate
    "LEARN_STEP": 128,  # Learning frequency
    "GAMMA": 0.99,  # Discount factor
    "GAE_LAMBDA": 0.95,  # Lambda for general advantage estimation
    "ACTION_STD_INIT": 0.6,  # Initial action standard deviation
    "CLIP_COEF": 0.2,  # Surrogate clipping coefficient
    "ENT_COEF": 0.01,  # Entropy coefficient
    "VF_COEF": 0.5,  # Value function coefficient
    "MAX_GRAD_NORM": 0.5,  # Maximum norm for gradient clipping
    "TARGET_KL": None,  # Target KL divergence threshold
    "UPDATE_EPOCHS": 4,  # Number of policy update epochs
    # Swap image channels dimension from last to first [H, W, C] -> [C, H, W]
    "CHANNELS_LAST": False,
}

num_envs = 16
env = make_vect_envs("LunarLander-v3", num_envs=num_envs)  # Create environment

observation_space = env.single_observation_space
action_space = env.single_action_space

pop = create_population(
    algo="PPO",  # RL algorithm
    observation_space=observation_space,  # State dimension
    action_space=action_space,  # Action dimension
    net_config=NET_CONFIG,  # Network configuration
    INIT_HP=INIT_HP,  # Initial hyperparameters
    population_size=INIT_HP["POP_SIZE"],  # Population size
    num_envs=num_envs,  # Number of vectorized envs
    device=device,
)

Training Loop

While off-policy RL algorithms can be considered more sample-efficient than on-policy algorithms, due to their ability to learn from experiences collected using a different or previous policy, on-policy algorithms often do better in practice due to the improved stability during training. Currently, AgileRL includes an evolvable implementation of Proximal Policy Optimisation (PPO). This algorithm can be used in a variety of settings and is widely popular across domains including robotics, games, finance, and RLHF.

The setup for PPO is very similar to the off-policy example above, except it does not require the use of an experience replay buffer. It also requires some different hyperparameters, shown below in the custom loop.

The easiest way to train a population of agents using PPO is to use our on-policy training function:

from agilerl.training.train_on_policy import train_on_policy

trained_pop, pop_fitnesses = train_on_policy(
    env=env,                              # Gym-style environment
    env_name="LunarLander-v3",  # Environment name
    pop=agent_pop,  # Population of agents
    max_steps=200000,  # Max number of training steps
    evo_steps=10000,  # Evolution frequency
    eval_steps=None,  # Number of steps in evaluation episode
    eval_loop=1,  # Number of evaluation episodes
    target=200.,  # Target score for early stopping
    tournament=tournament,  # Tournament selection object
    mutation=mutations,  # Mutations object
    wb=True,  # Weights and Biases tracking
)

Note

Known Gymnasium issue - running vectorize environments as top-level code (without if __name__ == "__main__":) may cause multiprocessing errors. To fix, run the above as a method under main, e.g.

def train_agent():
    # ... training code

if __name__ == "__main__":
    train_agent()

Alternatively, use a custom on-policy training loop:

Example Custom Training Loop

import numpy as np
import torch
from tqdm import trange
from agilerl.hpo.mutation import Mutations
from agilerl.hpo.tournament import TournamentSelection
from agilerl.utils.utils import create_population, make_vect_envs, default_progress_bar

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

NET_CONFIG = {
    "encoder_config": {
        "hidden_size": [32, 32], # Encoder hidden size
        "activation": "ReLU"
        },
    "head_config": {
        "hidden_size": [32]  # Head hidden size
    }
}

INIT_HP = {
    "POP_SIZE": 6,  # Population size
    "BATCH_SIZE": 128,  # Batch size
    "LR": 1e-3,  # Learning rate
    "LEARN_STEP": 128,  # Learning frequency
    "GAMMA": 0.99,  # Discount factor
    "GAE_LAMBDA": 0.95,  # Lambda for general advantage estimation
    "ACTION_STD_INIT": 0.6,  # Initial action standard deviation
    "CLIP_COEF": 0.2,  # Surrogate clipping coefficient
    "ENT_COEF": 0.01,  # Entropy coefficient
    "VF_COEF": 0.5,  # Value function coefficient
    "MAX_GRAD_NORM": 0.5,  # Maximum norm for gradient clipping
    "TARGET_KL": None,  # Target KL divergence threshold
    "UPDATE_EPOCHS": 4,  # Number of policy update epochs
}

num_envs = 16
env = make_vect_envs("LunarLander-v3", num_envs=num_envs)  # Create environment
observation_space = env.single_observation_space
action_space = env.single_action_space

# RL hyperparameters configuration for mutation during training
hp_config = HyperparameterConfig(
    lr = RLParameter(min=1e-4, max=1e-2),
    batch_size = RLParameter(min=8, max=1024),
    learn_step = RLParameter(min=64, max=1024)
)

pop = create_population(
    algo="PPO",  # RL algorithm
    observation_space=observation_space,  # State dimension
    action_space=action_space,  # Action dimension
    net_config=NET_CONFIG,  # Network configuration
    INIT_HP=INIT_HP,  # Initial hyperparameters
    hp_config=hp_config,  # Hyperparameters configuration
    population_size=INIT_HP["POP_SIZE"],  # Population size
    num_envs=num_envs,  # Number of vectorized envs
    device=device,
)

tournament = TournamentSelection(
    tournament_size=2,  # Tournament selection size
    elitism=True,  # Elitism in tournament selection
    population_size=INIT_HP["POP_SIZE"],  # Population size
    eval_loop=1,  # Evaluate using last N fitness scores
)

mutations = Mutations(
    no_mutation=0.4,  # No mutation
    architecture=0.2,  # Architecture mutation
    new_layer_prob=0.2,  # New layer mutation
    parameters=0.2,  # Network parameters mutation
    activation=0,  # Activation layer mutation
    rl_hp=0.2,  # Learning HP mutation
    mutation_sd=0.1,  # Mutation strength  # Network architecture
    rand_seed=1,  # Random seed
    device=device,
)

max_steps = 200000  # Max steps
evo_steps = 10000  # Evolution frequency
eval_steps = None  # Evaluation steps per episode - go until done
eval_loop = 1  # Number of evaluation episodes
total_steps = 0

# TRAINING LOOP
print("Training...")
pbar = default_progress_bar(max_steps)
while np.less([agent.steps[-1] for agent in pop], max_steps).all():
    pop_episode_scores = []
    for agent in pop:  # Loop through population
        agent.set_training_mode(True)

        obs, info = env.reset()  # Reset environment at start of episode
        scores = np.zeros(num_envs)
        completed_episode_scores = []
        steps = 0

        for _ in range(-(evo_steps // -agent.learn_step)):

            observations = []
            actions = []
            log_probs = []
            rewards = []
            dones = []
            values = []

            done = np.zeros(num_envs)

            learn_steps = 0
            for idx_step in range(-(agent.learn_step // -num_envs)):
                # Get next action from agent
                action, log_prob, _, value = agent.get_action(obs)

                # Clip to action space
                if isinstance(agent.action_space, spaces.Box):
                    if agent.actor.squash_output:
                        clipped_action = agent.actor.scale_action(action)
                    else:
                        clipped_action = np.clip(action, agent.action_space.low, agent.action_space.high)
                else:
                    clipped_action = action

                # Act in environment
                next_obs, reward, terminated, truncated, info = env.step(clipped_action)
                next_done = np.logical_or(terminated, truncated).astype(np.int8)

                total_steps += num_envs
                steps += num_envs
                learn_steps += num_envs

                observations.append(obs)
                actions.append(action)
                log_probs.append(log_prob)
                rewards.append(reward)
                dones.append(dones)
                values.append(value)
                obs = next_obs
                done = next_done
                scores += np.array(reward)

                for idx, (d, t) in enumerate(zip(terminated, truncated)):
                    if d or t:
                        completed_episode_scores.append(scores[idx])
                        agent.scores.append(scores[idx])
                        scores[idx] = 0

            pbar.update(learn_steps // len(pop))

            experiences = (
                observations,
                actions,
                log_probs,
                rewards,
                dones,
                values,
                next_obs,
                next_done,
            )
            # Learn according to agent's RL algorithm
            agent.learn(experiences)

        agent.steps[-1] += steps
        pop_episode_scores.append(completed_episode_scores)

    # Evaluate population
    fitnesses = [
        agent.test(
            env,
            max_steps=eval_steps,
            loop=eval_loop,
        )
        for agent in pop
    ]
    mean_scores = [
        (
            np.mean(episode_scores)
            if len(episode_scores) > 0
            else "0 completed episodes"
        )
        for episode_scores in pop_episode_scores
    ]

    pbar.write(
        f"--- Global steps {total_steps} ---\n"
        f"Steps: {[agent.steps[-1] for agent in pop]}\n"
        f"Scores: {mean_scores}\n"
        f"Fitnesses: {['%.2f' % fitness for fitness in fitnesses]}\n"
        f"5 fitness avgs: {['%.2f' % np.mean(agent.fitness[-5:]) for agent in pop]}\n"
    )

    # Tournament selection and population mutation
    elite, pop = tournament.select(pop)
    pop = mutations.mutation(pop)

    # Update step counter
    for agent in pop:
        agent.steps.append(agent.steps[-1])

pbar.close()
env.close()

Training Loop for Recurrent On-Policy Algorithms

Recurrent on-policy algorithms require a different training loop to the standard on-policy algorithms. This is because the agent needs to maintain a hidden state between steps, which is not possible with the standard training loop. AgileRL currently supports recurrent policies to be used with PPO. To use a recurrent policy, users must set both recurrent and use_rollout_buffer to True as arguments to the algorithm.

Example Population Creation

import torch
from agilerl.rollouts.on_policy import collect_rollouts_recurrent
from agilerl.utils.utils import create_population, make_vect_envs, default_progress_bar

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

NET_CONFIG = {
    "encoder_config": {
        "hidden_state_size": 64,
        "num_layers": 1,
        "max_seq_len": 512,
    },
    "head_config": {
        "hidden_size": [64],
    }
}

INIT_HP = {
    "POP_SIZE": 6,  # Population size
    "BATCH_SIZE": 128,  # Batch size
    "LR": 1e-3,  # Learning rate
    "LEARN_STEP": 128,  # Learning frequency
    "GAMMA": 0.99,  # Discount factor
    "GAE_LAMBDA": 0.95,  # Lambda for general advantage estimation
    "ACTION_STD_INIT": 0.6,  # Initial action standard deviation
    "CLIP_COEF": 0.2,  # Surrogate clipping coefficient
    "ENT_COEF": 0.01,  # Entropy coefficient
    "VF_COEF": 0.5,  # Value function coefficient
    "MAX_GRAD_NORM": 0.5,  # Maximum norm for gradient clipping
    "RECURRENT": True,
    "USE_ROLLOUT_BUFFER": True,
    "TARGET_KL": None,  # Target KL divergence threshold
    "UPDATE_EPOCHS": 4,  # Number of policy update epochs
}

# Create environment
num_envs = 16
env = make_vect_envs("LunarLander-v3", num_envs=num_envs)

observation_space = env.single_observation_space
action_space = env.single_action_space

# Create population
pop = create_population(
    algo="PPO",  # RL algorithm
    observation_space=observation_space,  # State dimension
    action_space=action_space,  # Action dimension
    net_config=NET_CONFIG,  # Network configuration
    INIT_HP=INIT_HP,  # Initial hyperparameters
    population_size=INIT_HP["POP_SIZE"],  # Population size
    num_envs=num_envs,  # Number of vectorized envs
    device=device,
)

tournament = TournamentSelection(
    tournament_size=2,  # Tournament selection size
    elitism=True,  # Elitism in tournament selection
    population_size=INIT_HP["POP_SIZE"],  # Population size
    eval_loop=1,  # Evaluate using last N fitness scores
)

mutations = Mutations(
    no_mutation=0.4,  # No mutation
    architecture=0.2,  # Architecture mutation
    new_layer_prob=0.2,  # New layer mutation
    parameters=0.2,  # Network parameters mutation
    activation=0,  # Activation layer mutation
    rl_hp=0.2,  # Learning HP mutation
    mutation_sd=0.1,  # Mutation strength  # Network architecture
    rand_seed=1,  # Random seed
    device=device,
)

max_steps = 200000  # Max steps
evo_steps = 10000  # Evolution frequency
eval_steps = None  # Evaluation steps per episode - go until done
eval_loop = 1  # Number of evaluation episodes
total_steps = 0

# TRAINING LOOP
print("Training...")
pbar = default_progress_bar(max_steps)
while np.less([agent.steps[-1] for agent in pop], max_steps).all():
    pop_episode_scores = []
    for agent in pop:  # Loop through population
        steps = 0
        completed_episodes = []
        last_obs, last_done, last_scores, last_info = None, None, None, None
        for _ in range(-(evo_steps // -agent.learn_step)):
            # Collect rollouts and save in buffer
            episode_scores, last_obs, last_done, last_scores, last_info = (
                collect_rollouts_recurrent(
                    agent,
                    env,
                    last_obs=last_obs,
                    last_done=last_done,
                    last_scores=last_scores,
                    last_info=last_info,
                )
            )

            agent.learn()  # Learn from rollout buffer

            # Update step counter and scores
            total_steps += agent.learn_step
            steps += agent.learn_step
            agent.steps[-1] += agent.learn_step
            completed_episodes += episode_scores

        pop_episode_scores.append(
            np.mean(completed_episodes)
            if len(completed_episodes) > 0
            else "0 completed episodes"
        )
        pbar.update(steps // len(pop))

    # Evaluate population
    fitnesses = [
        agent.test(
            env,
            max_steps=eval_steps,
            loop=eval_loop,
        )
        for agent in pop
    ]

    pbar.write(
        f"--- Global steps {total_steps} ---\n"
        f"Steps: {[agent.steps[-1] for agent in pop]}\n"
        f"Scores: {pop_episode_scores}\n"
        f"Fitnesses: {['%.2f' % fitness for fitness in fitnesses]}\n"
        f"5 fitness avgs: {['%.2f' % np.mean(agent.fitness[-5:]) for agent in pop]}\n"
    )

    if any(score >= required_score for score in pop_episode_scores):
        print(
            f"\nAgent achieved required score {required_score}. Stopping training."
        )
        elite, _ = tournament.select(pop)
        break

    # Tournament selection and population mutation
    elite, pop = tournament.select(pop)
    pop = mutations.mutation(pop)

    # Update step counter
    for agent in pop:
        agent.steps.append(agent.steps[-1])

pbar.close()
env.close()