Modular Neural Networks for Model-Free Behavioral Learning

Abstract

Future robots/agents will need to perform situation specific behaviors for each user. To cope with diverse and unexpected situations, model-free behavioral learning is required. We have constructed a modular neural network model based on reinforcement learning and demonstrated that the model can learn multiple kinds of state transitions with the same architectures and parameter values, and without pre-designed models of environments. We recently developed a modular neural network model equipped with a modified on-line modular learning algorithm, which is more suitable for neural networks and more efficient in learning. This paper describes the performances of constructed models using the probabilistically fluctuated Markov decision process including partially observable conditions. In the test transitions, the observed states probabilistically fluctuated. The new learning model is able to function in those complex transitions without specific adjustments for each transition.