agentdb-learn

# Learn Train an RL agent on episode data to derive a policy. ## When to use - Repeated task type with measurable reward and 50+ episodes - User asks to "train", "build a policy", "make it learn" - Pre-deployment of an autonomous agent that should pick actions itself ## Algorithms (the bandit picks the right one) | Algo | Best for | |---|---| | **Q-Learning** | Tabular state-spaces, discrete actions | | **SARSA** | On-policy variant, conservative exploration | | **DQN** | High-dimensional state, neural Q-fn | | **PPO** | Continuous control, high-dim action | | **Actor-Critic** | Baseline reduction, stable training | | **Policy Gradient** | Direct policy parameterization | | **Decision Transformer** | Offline RL on trajectories | | **MCTS** | Tree-search planning under known dynamics | | **Model-Based RL** | Sample-efficient when env model is learnable | If you don't know which to pick, call `agentdb_learning_route` first — the bandit suggests one based on past performance on similar task signatures. ## API ``` agentdb_learning_train( algorithm: <one of above> // or 'auto' to let bandit pick episodes: [<episodeId>, ...] // or task name → fetch automatically hyperparams: { lr, gamma, epsilon, ... } iterations: N ) ``` ## After training 1. The bandit logs the resulting reward distribution on the chosen algorithm. 2. The trained model is stored in the AgentDB skill library as a callable skill. 3. Future calls to `agentdb_learning_route(task)` may pick this trained skill if it scores well. ## Don't - Don't train on < 50 episodes — high-variance models overfit. - Don't train multiple algorithms in parallel "to see which wins" — that's the bandit's job, and parallel training pollutes the reward signal. - Don't ignore the route output. If `agentdb_learning_route` says "no algorithm has > 0.6 expected reward on this task", the answer is to gather more episodes, not to force-train.