Stochastic training of a biologically plausible spino-neuromuscular system model

A primary goal of evolutionary robotics is to create systems that are as robust and adaptive as the human body. Moving toward this goal often involves training control systems that process sensory information in a way similar to humans. Artificial neural networks have been an increasingly popular option for this because they consist of processing units that approximate the synaptic activity of biological signal processing units, i.e. neurons. In this paper we train a nonlinear recurrent spino-neuromuscular system (SNMS) model and compare the performance of genetic algorithms (GA)s, particle swarm optimizers (PSO)s, and GA/PSO hybrids. Several key features of the SNMS model have previously been modeled individually but have not been combined into a single model as is done here. The results show that each algorithm produces fit solutions and generates fundamental biological behaviors, such as tonic tension behaviors and triceps activation patterns, that are not explicitly trained. Categ...

• Type: Conference
• Year: 2007
• Where: GECCO
• Authors: Stanley Phillips Gotshall, Terence Soule
• Keywords: GECCO 2007, Neural Networks, Optimization, Particle Swarm Optimizers, Swarm Optimizers