The Immune System, Adaptation, and Machine Learning

Abstract:

The immune system is capable of learning, memory, and pattern recognition. By employinggenetic operators on a time scale fast enough to observeexperimentally,the immune systemis able to recognizenovel shapes without preprogramming. Here we describe a dynamical model for the immune system that is based on the network hypothesis of Jerne, and is simple enough to simulate on a computer. This model has a strong similarity to an approach to learning and artificialintelligence introduced by Holland, calledthe classifiersystem.We demonstrate that simpleversionsof the classifiersystemcan be cast as a nonlinear dynamical system, and explore the analogy between the immune and classifiersystems in detail. Through this comparison we hope to gain insight into the way they perform specifictasks, and to suggestnew approaches that might be of value in learning systems.