In task control., evolutionary optimization tends to favor controllers that solve the easier task instances but that fail to solve the harder ones. We call this the problem of hard instances. The doping-driven evolutionary control algorithm (DECA) is introduced to deal with the problem. The effectiveness of DECA is assessed on two task-control problems: a box-pushing task and a food-gathering task. The experimental results show DECA to generate controllers that can solve both the easy and hard instances of both task-control problems. We discuss the results by offering a qualitative explanation for DECA's success and comparing it to related techniques. We conclude that the problem of hard instances is alleviated by the application of DECA.
|Number of pages||29|
|Journal||Applied Artificial Intelligence|
|Publication status||Published - Mar 2008|