Evolving autonomous agent controllers as analytical mathematical models

Abstract

A novel Artificial Life paradigm is proposed where autonomous agents are controlled via geneticallyencoded Evolvable Mathematical Models (EMMs). Agent/environment inputs are mapped to agent outputs via equation trees which are evolved using Genetic Programming. Equations use only the four basic mathematical operators: addition, subtraction, multiplication and division. Experiments on the discrete Double-T Maze with Homing problem are performed; the source code has been made available. Results demonstrate that autonomous controllers with learning capabilities can be evolved as analytical mathematical models of behavior, and that neuroplasticity and neuromodulation can emerge within this paradigm without having these special functionalities specified a priori.