using-deep-rl

Value-based and policy-gradient methods suit different problem structures. Value-based methods work well for discrete action spaces with well-defined state values; policy-gradient methods handle continuous actions and sparse rewards better. This router identifies your problem type—action space, data regime, and reward structure—then directs you to the specialized skill that matches your RL task.

Online RL learns from interactions the agent generates in real time; offline RL learns from a fixed dataset without further exploration. The router assesses your data regime and available resources, then routes you to online RL skills for interactive training or offline RL skills when you have pre-collected trajectories and cannot afford live interaction.

Discrete action spaces suit problems with a fixed set of choices (e.g., game moves); continuous spaces handle infinite action ranges (e.g., robot joint angles). The router identifies your action space type and problem constraints, then directs you to the appropriate algorithm family—discrete-focused or continuous-capable methods—within the deep RL pack.

Model-free RL learns value or policy functions directly from experience; model-based RL learns an environment model first, enabling planning. The router evaluates your problem's sample efficiency needs, computational budget, and whether an environment model is available, then routes you to model-free or model-based specialized skills accordingly.

Multi-agent RL handles coordination, competition, or mixed incentives across agents. The router recognizes multi-agent problem structure and directs you to MARL (multi-agent reinforcement learning) specialized skills, distinguishing them from single-agent methods and addressing agent interaction challenges.

Debugging RL failures requires systematic diagnosis of reward signals, exploration behavior, environment design, and algorithm parameters. The router identifies debugging as your core need and directs you to the RL debugging skill, which covers diagnosis workflows separate from algorithm selection and training.