Deep Reinforcement Learning Approach for Flocking Control Problem of Multi-Agents in Obstruction Environment
- Resource Type
- Conference
- Authors
- Meng, Li; Cheng, Jin; Zhang, Han; Dong, Yujiao
- Source
- 2023 42nd Chinese Control Conference (CCC) Chinese Control Conference (CCC), 2023 42nd. :1-6 Jul, 2023
- Subject
- Computing and Processing
Robotics and Control Systems
Signal Processing and Analysis
Transportation
Deep learning
Potential energy
Training
Simulation
Design methodology
Neural networks
Fitting
flocking control
multi-agents deterministic strategy gradient
improved artificial potential energy function
explicit communication mechanism
- Language
- ISSN
- 1934-1768
A deterministic strategy gradient algorithm is proposed to train the behaviour of agents for the flocking control problem of multi-agents. Flocking behaviour and obstacle avoidance behaviour are learned by adding an improved artificial potential energy function as a reward mechanism. Furthermore, an explicit communication mechanism is established between multi-agents, which increases the input information of critics' network, so that it can evaluate the actions of agents more effectively. The flocking property and obstacle avoidance behaviour of multi-agent in obstacle environment are realized with the proposed method. Simulation results show that flocking control under explicit communication condition can effectively shorten the training time and achieve more stable control effect.