| LMAS: Adaptive Learning for Multi-Agent Coordination and Control |
|
Organization responsible: Delft University of Technology People involved:
Project description: In many control problems, the system to be controlled is insufficiently understood or too complex to design a solution in advance. Instead, a solution (control law) has to be learned. In tasks that are changing over time, pre-designed solutions also become inappropriate, and a good solution must adapt to the changes. Consider, for instance, a robot part of a multi-robot team, that has to perform some task in an unknown environment. Because the environment is unknown, a solution cannot be designed in advance. Because the other agents are making changes to the environment, a fixed solution is inappropriate. Even if a mathematical model of a complex task is known in advance, it might be very challenging to pre-program a good control law. Reinforcement learning promises a general solution to these problems. In theory, a reinforcement learner can learn an optimal control law from (simulated or real) interaction with a nonlinear, stochastic system, without any prior knowledge about the solution. An optimal control law dictates what actions to take in every state in order to maximize the performance. Performance is measured in RL by a scalar reward signal. However, the practical application of RL suffers from important difficulties. The main problem is that classical RL algorithms only work when the state-action space of the problem has a finite (and not too large) number of elements. Therefore, approximate algorithms are necessary in practice, where state-action spaces are usually large or continuous. In this project we therefore develop effective approximate reinforcement learning algorithms for continuous-variable tasks. Publications: |