Understanding the Agent’s Brain: A Quantitative Approach

Abstract

In recent years there have been many works describing successful autonomous agents controlled by Evolved Artificial Neural Networks. Understanding the structure and function of these neurocon-trollers is important both from an engineering perspective and from the standpoint of the theory of Neural Networks. Here, we introduce a novel algorithm, termed PPA (Performance Prediction Algorithm), that quantitatively measures the contributions of elements of a neural system to the tasks it performs. The algorithm identifies the elements which participate in a behavioral task, given data about performance decrease due to knocking out (lesioning) sets of elements. It also allows the accurate prediction of performance due to multi-element lesions. The effectiveness of the new algorithm is demonstrated in two recurrent neural networks with complex interactions among the elements. The generality and scalability of this method make it an important tool for the study and analysis of evolved neurocontrollers.