Effective, quantitative, obscured observation-based faultdetection in multi-agent systems
Michael Lindner, Noa Agmon · Adaptive Agents and Multi-Agents Systems · 2014
A key challenge in multi agent systems is verifying that all agents obey the system rules at any moment [10, 6]. In behavior-based systems, each agent is in some ‘state’ at any moment. The system designer defines what states each agent might take, according to its teammates states (plans [9, 1] or policies [8]). Being a distributed system, no single agent has knowledge of the others. Monitoring for fault detection must rely on gathering information by communication or observation. This information is not always accurate [9, 5]. There are various types of approaches for that issue. Some researches define a-priory the possible failures, and identify them at run time [7, 4]. Others, which we adopt, prefer to define the allowed behavior, and identify exceptions from it [6, 8], in the following way: (a) Define a policy of the state-combinations agents are allowed to take; (b) at run time, each agent observes its teammate, deducing their possible states; (c) then, it compares them to those allowed by the policy and see if: (1) all the possible states are allowed (no fault); (2) none of them is allowed (fault); (3) some are allowed and some are not (possible fault). Since the overall number of joint states in the system is m (number of agents powered by the number of states each agent might take), the naive comparison has exponential complexity in both space and time. Various researches suggests ways for reducing that. In [8], a binary matrix based policy is used. Each matrix (supercombination or s-comb) represents the states each agent is allowed to take; a pol-