What problem does it solve? Training reinforcement learning agents involves many error-prone details: choosing the right algorithm, building valid Gymnasium environments, configuring vectorized training, and setting up evaluation and checkpointing. This Skill provides tested templates and reference guides that prevent common mistakes like invalid observation spaces, misconfigured callbacks, and incorrect VecEnv usage. ## Core Features & Use Cases - Algorithm Selection Guidance: Detailed comparison of PPO, SAC, TD3, DQN, A2C, DDPG, HER, and RecurrentPPO with hyperparameter recommendations and a decision tree for picking the right algorithm. - Ready-to-Use Templates: Scripts for training agents with evaluation and checkpoint callbacks, evaluating trained models with video recording, and building custom Gymnasium environments validated with SB3's env_checker. - Vectorized Training & Callbacks: In-depth references on DummyVecEnv vs SubprocVecEnv, VecNormalize, VecFrameStack, and custom callback patterns for early stopping, learning rate scheduling, and TensorBoard logging. - Use Case: A robotics researcher needs to train a SAC agent on a custom continuous-control environment. They adapt the custom environment template, validate it with check_env, then train with the training script using gradient_steps=-1 and EvalCallback to save the best model. ## Quick Start Train a PPO agent on CartPole-v1 using the training template with four parallel environments and evaluation callbacks.
