Reinforcement Learning

knowledge skill

Trains a policy model by iteratively sampling actions, receiving rewards, and updating the model to maximize cumulative reward. Produces a trained policy artifact that maps states to actions.

Worked examples

• {"input": {"environment": "CartPole-v1", "algorithm": "PPO", "hyperparameters": {"learning_rate": 0.001, "batch_size": 64, "gamma": 0.99}}, "output": {"policy_model": "/models/ppo_cartpole_20240515.pth", "training_metrics": {"reward_mean": 495.0, "episode_count": 10000}, "config": {"learning_rate": 0.001, "batch_size": 64, "gamma": 0.99}}}
• {"input": {"environment": "Pendulum-v1", "algorithm": "DDPG", "hyperparameters": {"actor_lr": 0.0001, "critic_lr": 0.001, "tau": 0.005}}, "output": {"policy_model": "/models/ddpg_pendulum_20240515.pt", "training_metrics": {"reward_mean": -150.2, "episode_count": 5000}, "config": {"actor_lr": 0.0001, "critic_lr": 0.001, "tau": 0.005}}}

Input

• environment: str: identifier of a registered simulation or environment
• algorithm: str: reinforcement learning algorithm name (e.g., 'PPO', 'DQN')
• hyperparameters: dict: algorithm-specific settings (e.g., learning_rate, batch_size)

Output

• policy_model: str: path to saved policy model file
• training_metrics: dict: metrics logged during training (e.g., {'reward_mean': 120.5, 'episode_count': 5000})
• config: dict: final hyperparameters used for training

Details

• Skill type: knowledge skill
• Safety level: safe_public_research
• Version: 1.0.0