Lifelong learning approach to intelligent agents modeling

Abstract

In this paper, we presented an application of neural network-based models in the intelligent agent domain. The use of neural networks has the advantage that the model can adapt itself to changing environment conditions by simple retraining steps. This is illustrated by two functions needed for the lifelong learning-based modeling of an intelligent robotic agent: The function for generalizing sensor observations to conceptual states, and the function for modeling the effects of robot actions on the conceptual state space. An example shows that the algorithms can be applied in real-world environments.