Evaluation & Reliability
Reward Hacking
Reward hacking occurs when an agent obtains a high reward by exploiting the measured objective while failing the intended goal. The problem is a mismatch between what is rewarded and what actually matters. It does not require human-like malicious intent, and improving the reported score can make the real outcome worse.
Reviewed
Sources
Free complete lesson
This definition and the complete practical lesson are free. The interactive reader loads the lesson examples without a membership.