Using neural networks for synthesizing importance sampler database in reinforcement learning