Evolved Agents: Predator/Prey Game: Multitask

Agents are evolved to solve two tasks: Acting as predators to attack a prey agent, and acting as prey to escape a predator. Multiobjective neuroevolution is used to accomplish these tasks. Multitask networks have a separate set of output neurons corresponding to each task, i.e. one output mode per task. Therefore, it should be easy for Multitask to perform well in both tasks. Instead, Multitask networks always perform great in the Prey task (escaping), so sometimes perform poorly in the Predator task (attacking).