Decentralized learning in pursuit-evasion differential games with multi-pursuer and single-superior evader

Abstract

In this paper, we consider a multipursuer single-superior-evader pursuit-evasion differential game where the speed of the evader is similar to the speed of each pursuer. A new fuzzy reinforcement learning algorithm is proposed in this work for this game. Each pursuer of the game uses the proposed algorithm to learn its control strategy. The proposed algorithm of each pursuer uses the residual gradient fuzzy actor critic learning (RGFACL) algorithm to tune the parameters of the fuzzy logic controller (FLC) of the pursuer. The proposed algorithm uses a formation control approach in the tuning mechanism of the FLC of the learning pursuer so that the learning pursuer or the other learning pursuers can capture the superior evader. The formation control mechanism used by the proposed algorithm guarantees that the pursuers are distributed around the superior evader in order to avoid collision between pursuers. The formation control mechanism also guarantees that the capture regions of each two adjacent pursuers overlap or at least border each other so that the capture of the superior evader will be guaranteed. The proposed algorithm is a decentralized algorithm as no communication among pursuers is required. The only information that the proposed algorithm of each learning pursuer requires is the position and the speed of the superior evader. The proposed algorithm is used to learn a multi-pursuer single-superior-evader pursuitevasion differential game. The simulation results show the effectiveness of the proposed algorithm as the superior evader is always captured by one or some of the pursuers learning their strategies by the proposed algorithm.

Learning in the Multi-Robot Pursuit Evasion Game

“…Furthermore, this thesis discusses the problem of multi-pursuer single-evader PE games in which all players have equal capabilities. Though this game format was previously investigated in [23] it was assumed that the speed of the evader was known, which makes the algorithm inappropriate for practical use. Therefore, this assumption not considered here, which provides motivation for proposing another learning algorithm that could teach a group of pursuers how to capture a single evader in a decentralized manner.…”

“…As mentioned earlier, although there are numerous papers discussing different methods, including learning methods to solve the problem of the PE game with slow evaders, there are only a few that address multi-player PE differential games with superior evaders [8,[23][24][25][132][133][134][135][136]. The authors in [8] proposed a sufficiency condition to capture superior evaders.…”

“…al [23,136] proposed two decentralized learning algorithms for the PE game problem with a single-superior evader. The first algorithm [23] enables a team of pursuers with equal speed to capture a single evader with a similar speed. This learning algorithm is based on the condition that was proposed in [8], as well as a specific formation control strategy.…”

“…However, when the distance between the evader and the nearest pursuer is less than a specific tolerance distance, the strategy of the evader is to run away from that pursuer. In [23,136], there is a necessity to calculate the capture angle to obtain the required control signal for each pursuer to capture the superior evader. This thesis proposes a new reward function formulation for the FACLA algorithm, that enables a group of pursuers to capture a single evader in a decentralized manner and without finding the capture angle.…”

On Multi-Agent Reinforcement Learning in Matrix, Stochastic and Differential Games

“…This work is the first piece of work that presents a decentralized learning algorithm to capture a superior evader in a pursuit-evasion differential game without using any type of discretization for the state or action spaces. The proposed work is published in [138].…”