PROJECT TITLE :

Evolving Multimodal Networks for Multitask Games

ABSTRACT:

Intelligent opponent behavior makes video games fascinating to human players. Evolutionary computation will discover such behavior, however, it's challenging to evolve behavior that consists of multiple separate tasks. This paper evaluates three ways of meeting this challenge via neuroevolution: 1) multinetwork learns separate controllers for every task, which are then combined manually; 2) multitask evolves separate output units for every task, but shares data at intervals the network's hidden layer; and 3) mode mutation evolves new output modes, and includes a means to arbitrate between them. Whereas the first 2 strategies need that the task division be known, mode mutation will not. Leads to Front/Back Ramming and Predator/Prey games show that each of those ways has totally different strengths. Multinetwork is good in each domains, cashing in on the clear division between tasks. Multitask performs well in Front/Back Ramming, in which the relative problem of the tasks is even, however poorly in Predator/Prey, in which it's lopsided. Interestingly, mode mutation adapts to this asymmetry and performs well in Predator/Prey. This result demonstrates how a person's-specified task division is not always the simplest. Altogether the results recommend how human information and learning will be combined most effectively to evolve multimodal behavior.