Exploring the Relationship Between Topology and Function in Evolved Neural Networks

• Published in: 2020 IEEE Symposium Series on Computational Intelligence (SSCI) pp. 2304 - 2311
• Main Authors: Showalter, Ian, Schwartz, Howard
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2020
• Subjects: Artificial Neural Network; Autonomous Vehicle; Biological neural networks; Complexity theory; Evolutionary computation; Games; Hebbian Learning; Lamarckian Inheritance; Multiobjective; Network topology; Neurocontrollers; Neuroevolution; Neuromodulation; Pursuit-Evasion; Topology; Unsupervised Learning
• DOI: 10.1109/SSCI47803.2020.9308196

Understanding the relationship between structure and function in neural networks is essential to explaining their operation. Greater awareness of the link between topology and application could lead to wider adoption, particularly in mission-critical systems. Here, we examine and analyze the topology of very small, minimally sized neurocontrollers that have been evolved for an extended number of generations. Previously demonstrated Lamarckian-inherited neuromodulated evolutionary neurocontrollers are synthesized to operate a simulated vehicle pursuing a basic evader vehicle in the pursuit-evasion game. Both vehicles are subject to the effects of mass and drag. Constraints in the number of neurons and synapses are used to control network size. Additional objectives are added to the multiobjective optimization algorithm to encourage the selection of neural networks with the fewest neurons and synapses. It is shown that patterns emerge in the neuromodulatory neurons, in the direct connections between neurocontroller inputs and outputs, and that topologies similar to those used in classical control are evolved. Additionally, a neurocontroller constructed from the most commonly occurring neurons that successfully capture the evader is demonstrated.