We present the the concept of Guided Learning, which out-lines a framework in which a reinforcement learning agent can effectively’ask for help’ as it encounters stagnation. Either a human or expert agentsupervisor can then effectively ’guide’ the agent as to how to progressbeyond the point of stagnation. This guidance is then encoded in a novelway using a separately trained neural network referred to as a ’TaughtResponse Memory’ that can be recalled when another ’similar’ situa-tion arises in the future. This paper applies Guided Learning on topof an evolutionary algorithm but also shows how Guided Learning isalgorithm independent and can be applied in any reinforcement learn-ing context. The results show that our initial implementation of GuidedLearning provided in this paper gives superior performance and yields,on average, an increase of 136% in the rate of progression of the mostfit genome with best and worst case results yielding 137% and 110%respectively and an average increase of 112% in rate of progression forthe average genome with best and worst case results of 558% and 47%respectively. All results were achieved with minimal guidance. Such re-sults occur because the agent can exploit more information and thus,the need for exploration of the solution space is reduced. The results ob-tained show good promise for Guided Learnings potential as such resultswere obtained with only a partial implementation and much future workstill remains.

beelgroup Goto Github PK

Intelligent Systems Group, University of Siegen's Projects

Recommend Projects

Recommend Topics

Recommend Org