Reward Function
A reward function is the rule that scores an AI agent's actions or outcomes in reinforcement learning, defining what the agent is trying to maximise.
What is Reward Function?
Designers use the reward function to express the goal they want the system to pursue. Writing a reward that captures the real goal is hard: if it misses something important, the agent may find ways to score highly that the designer did not intend, which is a central concern in AI safety.
How it works
After each action or episode the environment returns a number. The learning algorithm adjusts the agent's behaviour to increase the total reward it expects to collect.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions