A reinforcement learning algorithm for a class of dynamical environments using neural networks

In many conventional approaches, when the environment is dynamically varied for agents, the models of agents are retrained in order to adapt to the current environment. However, when the same environments reappear in the future, it is not efficient to discard or modify the current model. To learn efficiently in this situation, we present new agent architecture. In this paper, we added extra models to the RAN-LTM agent model so that it can work well under a class of dynamic environments. In order to adapt rapidly to dynamic environments, it might be natural to consider that agents possess capability to store only essential knowledge, capability to retrieve proper knowledge, capability to detect environmental changes accurately.

[1]  Sebastian Thrun,et al.  Lifelong robot learning , 1993, Robotics Auton. Syst..

[2]  John C. Platt.  A Resource-Allocating Network for Function Interpolation , 1991, Neural Computation.

[3]  Andrew McCallum,et al.  Instance-Based Utile Distinctions for Reinforcement Learning with Hidden State , 1995, ICML.

[4]  Marco Wiering,et al.  Reinforcement Learning in Dynamic Environments using Instantiated Information , 2001, ICML.

[5]  Richard S. Sutton,et al.  Reinforcement Learning: An Introduction , 1998, IEEE Trans. Neural Networks.

[6]  Andrew W. Moore,et al.  Locally Weighted Learning , 1997, Artificial Intelligence Review.

[7]  Shigeo Abe,et al.  A reinforcement learning algorithm for neural networks with incremental learning ability , 2002, Proceedings of the 9th International Conference on Neural Information Processing, 2002. ICONIP '02..

[8]  Andrew W. Moore,et al.  Locally Weighted Learning for Control , 1997, Artificial Intelligence Review.