Parallel exploration via negatively correlated search

Abstract Effective exploration is key to a successful search process. The recently proposed negatively correlated (NCS) tries achieve this by coordinated parallel exploration, where set of processes are driven be so that different promising areas the space can visited simultaneously. Despite applications NCS, behaviors were mostly devised intuition, while deeper (e.g., mathematical) understanding missing. In paper, more principled namely NCNES, presented, showing equivalent process seeking probabilistic models both lead solutions high quality and distant from previous obtained models. Reinforcement learning, for which particular importance, considered empirical assessment. NCNES applied directly train deep convolution network with 1.7 million connection weights playing Atari games. Empirical results show significant advantages especially on games uncertain delayed rewards, highly owed effective ability.