DISTRIBUTED REINFORCEMENT LEARNING

In multi-agent systems two forms of learning can be distinguished: centralized learning, that is, learning done by a single agent independent of the other agents; and distributed learning, that is, learning that becomes possible only because several agents are present. Whereas centralized learning has been intensively studied in the field of artificial intelligence, distributed learning has been completely neglected until a few years ago. This paper summarizes work done on distributed reinforcement learning. The problem addressed is how multiple agents can learn to coordinate their actions such that they collectively solve a given environmental task. Two learning algorithms called ACE and DFG are described that provide answers to the following two questions: How can multiple agents learn which actions have to be carried out concurrently? How can multiple agents learn which sets of concurrent actions have to be carried out sequentially? Initial experimental results are provided which illustrate the learning abilities of these algorithms.