First published 2 July 2012
Analysis of Evolved Agents Performing Referential Communication
A pair of Continuous-time Recurrent Neural Network (CTRNN) based agents called "Sender" and "Receiver" is evolved on a circular world. Their collective objective is to communicate and move to a target — the Sender needs to communicate the address of a target location on the circle, and the Receiver needs to move to that location after receiving the communication. In extension of previous work (Williams and Beer, 2008), the agents are evolved under conditions different from the original work. Qualitative analysis of the most successful agent-pair shows that the Receiver's behavior is reminiscent of Newton's equations of motion in relating its initial velocity to the target address communicated to it. Further analysis using information-theoretic tools reveals a pair of neurons that hold crucial information required for the successful functioning of the Receiver. They are also shown to employ the same kind of information for slightly different purposes.