Behavioral diversity with multiple behavioral distances

摘要: Recent results in evolutionary robotics show that explicitly encouraging the behavioral diversity of candidate solutions drastically improves convergence many experiments. The performance this technique depends, however, on choice a similarity measure (BSM). Here we propose experimenter does not actually need to choose: provided several measures are conceivable, using them all could lead better than choosing single one. Values computed by BSM can be averaged, which is computationally expensive because it requires computation at each generation, or randomly switched user-chosen frequency, cheaper alternative. We compare these two approaches experimental setups - ball collecting task and hexapod locomotion with five different BSMs. Results (1) run increases while avoiding choose most appropriate (2) switching between BSMs leads taking mean diversity, requiring less computational power.