Impacts of Mathematical Optimizations on Reinforcement Learning Policy Performance

Deep neural networks (DNN) now outperform competing methods in many academic and industrial domains. These high-capacity universal function approximators have recently been leveraged by deep reinforcement learning (RL) algorithms to obtain impressive results for many control and decision making problems. During the past three years, research toward pruning, quantization, and compression of DNNs has reduced the mathematical, and therefore time and energy, requirements of DNN-based inference. For example, DNN optimization techniques have been developed which reduce storage requirements of VGG-16 from 552MB to 11.3MB, while maintaining the full-model accuracy for image classification. Building from DNN optimization results, the computer architecture community is taking increasing interest in exploring DNN hardware accelerator designs. Based on recent deep RL performance, we expect hardware designers to begin considering architectures appropriate for accelerating these algorithms too. However, it is currently unknown how, when, or if the “noise” introduced by DNN optimization techniques will degrade deep RL performance. This work measures these impacts, using standard OpenAI Gym benchmarks. Our results show that mathematically optimized RL policies can perform equally to full-precision RL, while requiring substantially less computation. We also observe that different optimizations are better suited than others for different problem domains. By beginning to understand the impacts of mathematical optimizations on RL policy performance, this work serves as a starting point toward the development of low power or high performance deep RL accelerators.