Research On The Surrounding Control Of Multi-agent Systems

In recent years,with the development of computer technology and network technology,multiagent technology has penetrated into every aspect of our life.As a hot research direction of multiagent technology,multi-agent co-regulation and production has produced many applications,such as UAV formation performance,joint operation and joint rescue of ground robots,etc.As a common problem in the coordinated control of multi-agent,surrounding control mainly refers to the realization of surrounding a specific target by multiple agents using distributed communication information or local information by designing effective control protocols.Aiming at the problem of multi-agent surrounding control in two-dimensional space,the main research points of this thesis are as follows:A control method based on adaptive RBF(Radial Basis Function)neural network is proposed for a first-order system with a single non-cooperative target and multiple surrounding agents,which enables surrounding agents to surround the target in a given way.Considering that most targets of surrounding control are cooperative,this thesis proposes enveloping control for non-cooperative targets with escape capability.In this thesis,we simulate the escape strategy of non-cooperative targets by constructing a repulsive potential field around surrounding agents,and the potential field function is unknown to surrounding agents.Aiming at the unknown artificial potential field,the adaptive RBF neural network is designed to fit the potential field action function.On this basis,an adaptive control method based on a given bounding potential point is proposed to realize the bounding control by making the bounding agent track the given bounding potential point.Based on the potential point control method,a virtual leader-based surrounding control method is proposed to shorten the convergence time of surrounding control.After completing the theoretical proof through the Lyapunov function,two sets of comparative simulations are given to demonstrate the effectiveness and superiority of the proposed adaptive surrounding control algorithm.An adaptive surrounding controller based on geometric center estimator is proposed to solve the surrounding control problem of multiple targets for a first-order system with multiple unknown dynamic targets and surrounding agents.Considering that the motion state of the target is known in most surrounding control problems,this thesis presents surrounding control problems for several unknown dynamic targets.Multiple targets in the system move according to a given dynamic equation and surround the agent to obtain some state information of the target which communicates with the agent.In this thesis,a distributed geometric center estimator is designed to estimate the geometric center of the target smoothly.On this basis,the nonlinear term in the target dynamic model is fitted by the neural network,and a reasonable enveloped control strategy is designed.Finally,the enveloping agent is distributed on the enveloping circle with the center of the target geometry as the center of the circle in a given way.After completing the theoretical proof through the Lyapunov function,two groups of comparative simulations are given to demonstrate the superiority of the proposed adaptive surrounding control algorithm based on geometric center.A surrounding control method based on deep reinforcement learning is proposed for second-order systems with multiple targets and multiple enveloping agents.Considering that most surrounding controls depend on mathematical models,this thesis proposes a control method based on reinforcement learning,which can solve the problems without relying on detailed mathematical models.Firstly,the traditional surrounding control problem is transformed into a Markov game,and two control problems are proposed according to the control tasks.The first is to surround the unknown dynamic target in an annular formation.The purpose of the control is to make the surrounding agent evenly distributed on the enveloping circle with the geometric center of the target as the center of the circle.The second is the problem of enveloping the escaping target.The purpose of the control is to make the enveloping agent shrink the target into the convex hull formed by enveloping agent.To solve the above two problems,this thesis first introduced MADDPG algorithm and applied it to the surrounding control.Secondly,on the basis of MADDPG algorithm,LSTM(long short-term memory)network and Attention mechanism were introduced to optimize the control effect,and AR-MADDPG(Attention Recurrent-MADDPG)algorithm was proposed and applied in surrounding control.Finally,the corresponding simulation experimental environment is built,and the MADDPG algorithm and AR-MADDPG algorithm are used to compare the simulation experiments for the above two problems respectively,to verify the feasibility and effectiveness of the algorithm.