Apprentissage par renforcement

The agent interacts with an environment and tries to improve a cumulative reward objective. Reward is a designed signal and may only imperfectly represent the behavior people want.

Points clés à retenir

• Actions affect future observations.
• Reward design can create unintended incentives.
• Constrain exploration and test beyond one environment.

The agent observes a state or observation, chooses an action according to a policy, and receives feedback. An action can change the situations encountered later, so the task differs from predicting independent labels. A sequence of interactions may form an episode, such as one game or one simulated journey. Exploration tries actions to learn about their consequences. Exploitation uses what the agent has learned to pursue reward. Real deployments must constrain exploration when mistakes can affect people, equipment, or budgets. A simulation is often useful, but success inside it may depend on assumptions that fail outside it. Delayed rewards create a credit-assignment problem: which earlier actions helped or harmed the outcome? Algorithms estimate values or directly improve policies using experience. The choice of observation, action space, reward, and time horizon can matter as much as the algorithm name. Test whether a policy exploits loopholes in the reward. If a support agent earns reward for closing tickets, it might close unresolved requests. Include direct checks of task completion and unacceptable outcomes. Compare performance across varied starting conditions and report the cost of collecting experience.

1. In an invented three-step episode, rewards are 2, 0, and 5. With discount factor 0.9, the return from the first step is 2 + 0.9 × 0 + 0.9² × 5.
2. The result is 6.05. With no discounting, the total would be 7.
3. Changing the reward definition or horizon could change which policy is preferred even in the same environment.

This arithmetic example explains the objective; it is not a measured reinforcement-learning result.