Decentralized Multi-agent Patrolling Strategies Using Global Idleness Estimation

Mehdi Othmani-Guibourg, Amal El Fallah Seghrouchni, Jean-Loup Farges

Abstract: This paper presents preliminary results in the challenge of developing decentralised strategies approaching the performances of centralised ones. Indeed, the latter is better than the former due to centralisation of information. The approach studied here involves the estimation of real node idlenesses (as known by the coordinator) from the individual ones retained by each agent. This relation between real and individual idlenesses is learnt using traces of execution of a centralised strategy by optimising an error criterion. The strategy thereupon, uses online the learnt relation and is assessed according to certain evaluation criteria. The results indicate that such a relation between perceived and real idlenesses is not a function, leading to large values of the fitting criterion. Finally, the assessment of the strategy shows that performances are good in terms of mean interval but unsatisfactory in terms of quadratic mean interval.