Learning Co-ordination Skills by Exploiting the inherent Hierarchy of the Task

Graduate

Traditional Robot Learning is extremely slow and tedious if the problem is complex and the state space large. This necessitates the development of new machine learning techniques which speed up learning considerably by taking advantage of the task structure and introducing a hierarchy in terms of state and action space. On the other hand, extending flat learning to Multi Agent tasks is extremely challenging due to scalability issues ( the state space is exponential in terms of the number of agents ) and the computational challenges involved. Also, interactions among agents at a low level might interfere with learning the overall task and co-ordination skills. The use of hierarchy in multi agent learning makes it possible to learn co-ordination skills at the level of abstract actions, thus taking care of the aforementioned problems. The MaxQ method of hierarchical learning lends itself to easy extension to the multi agent case. The inherent structure of the task is captured in the hierarchy and the action space at the highest level of abstraction alone is used for learning co-ordination skills. This technique makes multi agent learning successful in large scale real life tasks. Results on a robot simulation task with two agents show that co-ordination skills can be learned efficiently using this method.

Return to Workshop Listing