SAC Pendulum

import gymnasium as gym

from rlforge.agents.policy_gradient import SACAgent
from rlforge.experiments import ExperimentRunner

num_envs = 8
envs = gym.make_vec("Pendulum-v1", num_envs=num_envs, vectorization_mode="async")

agent = SACAgent(
    state_dim=envs.observation_space.shape[1],
    action_dim=envs.action_space.shape[1],
    policy_net_architecture=(64, 64),
    q_net_architecture=(64, 64),
    actor_lr=3e-3,
    critic_lr=3e-3,
    alpha_lr=3e-3,
    discount=0.99,
    tau=0.005,
    update_frequency=10,
    buffer_size=1000000,
    mini_batch_size=256,
    update_start_size=256,
    tanh_squash=True,
    action_low=envs.action_space.low[0],
    action_high=envs.action_space.high[0],
    target_entropy_factor=0.9,
    device="cpu"
)

runner = ExperimentRunner(envs, agent)

results = runner.run_episodic_batch(
    num_runs=5,
    num_episodes=1500,
    max_steps_per_episode=None
)

rewards = results["rewards"]

runner.summary(last_n=20)

============================================================
 Experiment Summary (Episodic)
============================================================
Runs: 5
Average runtime per run: 107.417 seconds
Episodes per run (Max): 1500
First episode mean reward: -1217.885
Last episode mean reward: -179.092
Overall mean reward: -546.502
Mean reward (last 20 episodes): -190.037
First episode mean steps: 200.0
Last episode mean steps: 201.0
Overall mean steps: 201.0
============================================================

runner.plot_results()