A k-NN Based Perception Scheme for Reinforcement Learning
A perception scheme for Reinforcement Learning (RL) is developed as a function approximator. The main motivation for the development of this scheme is the need for generalization when the problem to be solved has continuous state variables. We propose a solution to the generalization problem in RL algorithms using a k-nearest-neighbor pattern classification (k-NN). By means of the k-NN technique we investigate the effect of collective decision making as a mechanism of perception and action-selection and a sort of back-propagation of its proportional influence in the action-selection process as the factor that moderate the learning of each decision making unit. A very well known problem is presented as a case study to illustrate the results of this k-NN based perception scheme.
KeywordsReinforcement Learning k-Nearest-Neighbors Collective Decision Making
Unable to display preview. Download preview PDF.
- 1.Reinforcement Learning. In: Sutton, R.S (ed.) SECS, vol. 173, Kluwer Academic Publishers (1992). Reprinted from of Machine Learning, vol. 8(3-4) (1992) Google Scholar
- 2.Sutton, R.S.: Reinforcement learning and artificial intelligence (2006)Google Scholar
- 3.Sutton, R.S.: Reinforcement learning architectures. In: Proc. Int. Symp. on Neural Information Processing, Kyushu Inst.of Technology, Japan (1992)Google Scholar
- 4.Sutton, R.S., Barto, A.G.: Reinforcement Learning. An Introduction. MIT Press, Cambridge (1998)Google Scholar
- 6.Dudani, S.A.: The distance-weighted k-nearest-neighbor rule. IEEE Transactions on Systems, Man and Cybernetics SMC-6(4), 325–327 (1976)Google Scholar