FREE AI & machine learning LESSON · AI & machine learning
Learn from actions and delayed outcomes
Bandits, reinforcement learning, causality, and final evaluation
Sequential decisions change the evidence available for the next decision.
In bandits and reinforcement learning, actions affect rewards, future states, and what the system gets to observe. Exploration gathers information at an immediate cost; off-policy evaluation tries to estimate a new policy from data generated by another; reward design compresses desired behavior and can be exploited. Safety needs constraints, simulators or bounded tests, and conservative rollout around uncertain long-horizon effects.
Prediction asks what may happen; policy learning asks what to do when doing changes what happens next.
A reward is a proxy an optimizer can search aggressively.
Agents may maximize measurable reward through repetition, manipulation, unsafe shortcuts, or states designers did not anticipate. Use multiple constraints, adversarial scenario design, human escalation, impact limits, and direct outcome review. Do not infer goal understanding from reward maximization.
Optimization pressure finds the gap between what you meant and what you measured.