Reward engineering translates business goals and rules into scalar feedback an agent can optimize. Since the agent responds to the number it receives, the reward must represent both desired outcomes and acceptable boundaries. section 2.6
Design the signal around the decisions the agent must make:
- When a task unfolds over time, consider stepwise rewards that give feedback at each decision rather than only at the end. Intermediate measures can approximate outcomes that take longer to observe. section 2.6.1
- Balance multiple goals and constraints so that a reward for one objective does not undermine the others. section 2.6
- Test and monitor the behavior the reward produces; an agent may exploit a loophole or optimize the wrong trade-off while still earning a high score. section 2.6.6 section 12.9.2
reward = ($0.02 * click_count) + ($0.10 * revenue) - ($0.05 * bounce_rate).
A weighted score combines clicks, revenue, and bounce rate, illustrating how a reward can encode competing signals. section 2.6.6
A risk penalty that overwhelms returns
In portfolio management, risk penalties set too high led an agent to move almost everything into cash and stop trading. The reward looked reasonable, but the learned behavior was useless; checking what the agent actually did revealed the problem. section 12.9.2
