Learn by Experience of Taking Actions

Some machines improve through trial and error, interacting with an environment to maximize rewards—a paradigm inspired by behavioral psychology. Unlike statistical learning, this involves sequential decision-making where actions influence future possibilities.

Imagine teaching a robot to play basketball without explicitly programming the rules. The robot starts by making random movements—some shots miss wildly, others accidentally score. Each time the ball goes through the hoop, the robot receives a "reward signal" that strengthens the neural connections that produced that successful action. Over thousands of attempts, the robot gradually discovers patterns: holding the ball this way, applying force at that angle, and adjusting for distance all increase its chances of success. The machine builds an internal model connecting actions to outcomes, becoming increasingly strategic about which moves to try next. What makes this approach powerful is that the machine discovers solutions we might never explicitly teach it, sometimes finding creative strategies that human experts hadn't considered.

Natural selection offers another powerful learning paradigm inspired by evolutionary biology. Genetic algorithms maintain populations of potential solutions that compete, with the fittest individuals surviving to reproduce. Each solution is encoded as a 'chromosome' representing parameters or rules, and solutions evolve through mechanisms like crossover (combining successful solutions) and mutation (introducing random variations). For example, when optimizing aerodynamic shapes, genetic algorithms might start with diverse random designs, evaluate their performance in simulated environments, and allow the best performers to contribute features to the next generation. Over many iterations, solutions naturally evolve toward optimality without explicit directions on how to improve. This approach excels at complex optimization problems with vast solution spaces, discovering novel solutions by exploring combinations a human designer might never consider.