6. Learning in multi-agent universes

It's not always possible to define completely and precisely, at the design stage, the environment in which an agent will evolve, nor the coordination rules of an ADM. It is also difficult to optimize agent behavior a priori, given the open, evolving and unpredictable nature of the environment and the complexity of the situations they may encounter. Machine learning enables an agent to gradually discover and adapt to its environment, to become more efficient or to learn how to coordinate with its associates to carry out a complex task. The example of a team of soccer robots is a case where learning enables a team to establish strategies, each teammate to learn to play a role (goalkeeper, defender, attacker, midfielder) in the service of the common goal (winning a match) and to adapt collective behavior to the opposing team, whose behavior and organization are not known a priori.