LazyAct: Lazy actor with dynamic state skip based on constrained MDP

Deep reinforcement learning has achieved significant success in complex decision-making tasks. However, the high computational cost of policies based on deep neural networks restricts their practical application. Specifically, each decision made by an agent requires a complete neural network computation, leading to a linear increase in computational cost with the number of interactions and agents. Inspired by human decision-making patterns, which involve reasoning only on critical states in continuous decision-making tasks without considering all states, we introduce the LazyAct algorithm. This algorithm significantly reduces the number of inferences while preserving the quality of the policy. Firstly, we incorporate a state skipping branch into the actor network to bypass states with minimal impact. Subsequently, we establish optimization objectives for single-agent and multi-agents inference, incorporating cost constraints based on the IMPALA and MAPPO frameworks, respectively. Finally, we utilize pre-training and fine-tuning techniques to train the policy network. Extensive experimental results indicate that LazyAct reduces the number of inferences by approximately 80% and 40% in single-agent and multi-agents scenarios, respectively, while sustaining comparable policy performance. The inferences reduction significantly decreases the time and FLOPs required by the LazyAct algorithm to complete tasks. Code is available here https://www.dropbox.com/scl/fo/wyoqo6q9gyt86zobfgbvx/h?\rlkey=0moyxsnoiisfs9y4h89hsou1l&dl=0.