Neural network compression for reinforcement learning tasks.

In real applications of Reinforcement Learning (RL), such as robotics, low latency, energy-efficient and high-throughput inference is very desired. The use of sparsity and pruning for optimizing Neural Network inference, and particularly to improve energy efficiency, latency and throughput, is a standard technique. In this work, we conduct a systematic investigation of the application of these optimization techniques with popular RL algorithms, specifically Deep Q-Network and Soft Actor Critic, in different RL environments, including MuJoCo and Atari, which yields up to a 400-fold reduction in the size of neural networks. This work presents a systematic study on the applicability limits of using pruning and quantization to optimize neural networks in RL tasks, with a perspective of deployment in hardware to reduce power consumption and latency, while increasing throughput.