Punctuated Anytime Learning with Fitness Biasing for Gait Generation
By: William Tarimo '12 and Moustafa Ndiaye '17
Advising Faculty: Gary Parker
Punctuated anytime learning (PAL) can be used to improve genetic evolution where evolutionary training takes place on a simple offline model with periodic checks on the actual robot. This technique provides a way of linking the actual robot with its simulation model and thus improving the output of the learning system. Fitness biasing, a type of PAL, modifies the internal workings of the evolutionary algorithm. A comparison of the performance of solutions on the actual robot and on the simulation model is used to bias the evolved solutions’ fitnesses towards better performance on the actual robot. In this work, we present an implementation of an actual hexapod training system that incorporates a cyclic genetic algorithm and PAL with fitness biasing.
Related Fields: Computer Science