¹Ericsson Research, Bangalore-560037, India.
²Ericsson Research, Kista-16480, Sweden.

---

Abstract—In large scale multi-agent systems, both planning for system goals and replanning during plan execution to handle failures are compute-intensive. Since replanning requires faster response time because it happens during plan execution, a lot of focus in the AI planning literature has been on incremental methods, such as plan repair, plan modification etc., which avoid re-synthesizing a complete plan. In this paper, we propose HIPR - an architecture and supporting algorithms for fast replanning in multi-agent systems with two fairly general characteristics: (1) where the agents are organized hierarchically based on attributes like location/administration and (2) where most failures are localized i.e. only a few agents are affected by the failure while most of the agents at large remain unaffected. HIPR exploits these characteristics to identify the smallest group of agents that are affected by the failure and to synthesize new plan fragments for only these agents. The localization to smaller number of agents generates smaller replanning problems and hence more efficient solutions. We illustrate application of HIPR on a small, yet realistic route planning use case.

Index Terms—Artificial intelligence, hierarchical domains, planning, replanning.

[PDF]

Cite: Swarup Kumar Mohalik, Mahesh Babu Jayaraman, Ramamurthy Badrinath, Aneta Vulgarakis Feljan, "HIPR: An Architecture for Iterative Plan Repair in Hierarchical Multi-agent Systems," Journal of Computers vol. 13, no. 3, pp. 351-359, 2018.