Online Planning for Stochastic Collaborative Privacy Preserving Planning

Oriel Uzan, Guy Shani

Keywords: multi-agent planning, privacy preserving planning, heuristics

Abstract: Collaborative multi-agent privacy preserving planning (CPPP) models problems where agents must work together to achieve joint goals, while keeping some information private. Recently, CPPP was extended to the stochastic case, where actions may fail, producing different effects than intended. Stochastic CPPP (SCPPP) problems can be solved using offline algorithms, such as RTDP. However, in many cases, we are not interested in computing a complete policy offline, and prefer to use an online approach, where one decides online on the next action only, without exploring the complete state space. This can allow us to scale to much larger problems. In this paper we thus explore online approaches for SCPPP. We suggest using a variant of the well known FF-Replan approach, adapted to CPPP, and a plan repair approach, where we try to locally return to the plan if an undesirable effect has occurred. We provide an empirical evaluation, comparing our approaches to an offline solver, showing that we can scale to much larger problems, and analyzing the strengths and weaknesses of our methods.