Reveal or Hide: Information Sharing in Multi-Agent Planning

Abstract

The ability to plan a sequence of action in order to achieve a given goal with respect to the initial conditions of the world is one of the crucial aspects of intelligence. It is no surprise, that this aspect has been thoroughly studied in the context of artificial intelligence since its very beginning. The same can be said about the study of multi-agent aspects of planning in the research field of multi-agent systems. Among the most important aspects of such multi-agent planning is information sharing, that is, which information should be shared by the agents and which not, and also how to share the information efficiently. We provide several perspectives on the issue of sharing or hiding information in multi-agent planning. We mostly focus on heuristic search with domain-independent heuristics which is a well-established approach both in classical and multi-agent planning. We advance the state of the art in a number of directions. Firstly, we focus on the distributed computation of heuristics. The main research question is how to achieve global heuristic guidance without explicitly communicating and revealing private parts of the planning problems respective to the particular agents. We approach this issue by providing a number of distributed variants of classical planning heuristics, both inadmissible and admissible (which are necessary for optimal planning). We use the acquired knowledge to design more general approaches for distributing relaxation heuristics and finally any heuristic (in an admissible way). We theoretically analyze the distributed heuristics (e.g., by showing their admissibility) and provide a thorough experimental evaluation, showing their superiority in speed or heuristic guidance compared to the same heuristics computed locally by the agents (that is, without sharing any information throughout the heuristic computation). Secondly, we propose a heuristic search algorithm which is able to balance the use of distributed and local heuristics. The distributed heuristic approach is not always the best choice. In many problems, the heuristic guidance of the locally computed heuristic is close to the distributed variant but without the computation and communication overheads. We solve the issue by allowing the search to use the local heuristic while computing the distributed heuristic and waiting for replies from other agents. This technique is able to balance the information sharing in most domains and problems and practically dominates each approach used separately. The resulting planner also improves on the state of the art in suboptimal multi-agent planning. Thirdly, we analyze information sharing in multi-agent planning in the context of privacy. In privacypreserving cooperative multi-agent planning, the agents want to cooperatively plan a sequence of actions but do not want to reveal their private knowledge. In realistic scenarios, avoiding explicit communication of the private information is not enough, the agents do not want to allow any other agent even to deduce such information from the communication protocol. The thesis builds on two major journal publications and a number of works published at the top-tier AI conferences. The designed algorithms are both theoretically analyzed and thoroughly experimentally evaluated. In order to allow for a more complete and rigorous comparison of existing multi-agent planners, we have co-organized the first Competition of Multi-Agent and Distributed Planners (CoDMAP) during the work on the above research topics. We have collaborated on the design of the formal domain and problem description language, we have designed the competition setup (in two tracks), implemented necessary software tools, and performed the evaluation. The description of the competition and the results relevant to other presented topics are provided as a part of the thesis.