Study and development of reinforcement learning multiagent collaborative algorithms for solving families of problems

Artificial intelligence algorithms enable autonomous agents to perform sophisticated tasks with great accuracy. A particularly relevant technique within artificial intelligence is that of reinforcement learning, which allows strategic planning and optimal control of an environment through interaction with it and without prior knowledge. Most current algorithms based on reinforcement learning allow us to learn the optimal policy that solves a specific problem, that is, the actions performed in each state that allow maximizing the reward obtained by the agent when it interacts with a given environment. It is of great interest the study of techniques that allow to obtain policies that are suitable not only for a specific problem, but for a whole family of similar problems. This new problem is called transfer learning, where the goal is that the knowledge obtained during the resolution of a problem is useful to solve other different but related problems. The goal of this final degree project is the development of distributed reinforcement learning algorithms in which multiple agents learn different instances of a family of problems, so that each agent learns by interacting with their particular environment, but the agents communicate with each other and cooperate to learn a policy that is good for the whole family of problems. We will also develop techniques to evaluate whether these algorithms generalize in some sense the solution to any similar problem