“School of Cognitive Sciences”
Back to Papers HomeBack to Papers of School of Cognitive Sciences
| Paper IPM / Cognitive Sciences / 8362 |
|
||||||
| Abstract: | |||||||
|
Multiagent credit assignment (MCA) is one of the major problems in the
realization of multiagent reinforcement learning. Since the environment usually is
not intelligent enough to qualify individual agents in a cooperative team, it is very
important to develop some methods for assigning individual agents' credits when
just a single team reinforcement is available. MCA cannot be solved in general
cases, using a single technique. Therefore, our goal in this research is first to
present a new view of the problem and second, to introduce a new idea of using
agents' knowledge to partially solve MCA. In this research, an approach that is
based on agents' learning histories and knowledge is proposed to solve the MCA
problem. Knowledge evaluation-based credit assignment (KEBCA) along with
certainty, a measure of agents' knowledge, is developed to judge agents' actions
and to assign them proper credits. The proposed KEBCA method is general,
however; we study it in some simulated extreme cases in order to gain a better
insight into MCA problem and to evaluate our approach in such cases. More
specifically, we study the effects of task type (and-type and or-type tasks) on
solving MCA problem in two cases. In the first case, in addition to the team
reinforcement, it is assumed that some extra information at the team level is
available. In the second case, such extra information does not exist. In addition,
performance of the system is examined in presence of some uncertainties in the
environment, modeled as noise on agents' actions. The information content of
team reinforcements and assumed extra information are theoretically calculated
and discussed. The mathematical calculations confirm the related simulation
results
Download TeX format |
|||||||
| back to top | |||||||
