On-Policy

Definitions

1. Describing reinforcement learning methods that learn from data generated by the current policy being improved rather than from a separate behavior policy. On-policy methods can be more stable in some settings but often less sample-efficient.

   In plain English: Learning from data generated by the current policy itself.

   Example: "The simulator training loop stayed on-policy so each update reflected the agent's current behavior rather than stale trajectories."

Related Terms

• AI Multi-Agent
• Alignment Tax
• Benchmark Saturation
• Capability Overhang
• Dialogue Policy
• Replay Buffer
• Safety Policy
• Trajectory