This week’s issue explores one of the central design problems in reinforcement learning: reward hacking.
In reinforcement learning, an agent learns by finding actions that earn more reward. But the reward is only our description of what success should look like. If that description leaves a loophole, the agent can achieve a high score while doing something useless—or even harmful.
The new interactive article builds the idea from the ground up through three small environments. You will see how a missing restriction, a measurement the agent can manipulate, or an outcome the objective forgot to count can each turn apparent success into failure—and learn how to recognize the same gaps elsewhere.