Q-Learning Algorithm Module in Hybrid Artificial Neural Network Systems

Abstract

Presented topic is from the research field called Artificial Life, but contributes also to the field of Artificial Intelligence (AI), Robotics and potentially into many other aspects of research. In this paper, there is reviewed and tested new approach to autonomous design of agent architectures. This novel approach is inspired by inherited modularity of biological brains. During designing of new brains, the evolution is not directly connecting individual neurons. Rather than that, it composes new brains by connecting larger, widely reused areas (modules). In this approach, agent architectures are represented as hybrid artificial neural networks composed of heterogeneous modules. Each module can implement different selected algorithm. Rather than describing this framework, this paper focuses on designing of one module. Such a module represents one component of hybrid neural network and can seamlessly integrate a selected algorithm into the node. The course of design of such a module is described on example of discrete reinforcement learning algorithm. The requirements posed by the framework are presented, the modifications on the classical version of algorithm are mentioned and then the resulting performance of module with expectations is evaluated. Finally, the future use cases of this module are described.