Optimizing and planning with queries for Communication in Ad Hoc Teamwork

Access full-text files

Date

2023-04-21

Authors

Macke, William

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

In recent years autonomous agents have become increasingly sophisticated and are able to handle more complex tasks compared to previous generations. There are many such tasks that due to their nature require or can be more effectively solved by multiple agents compared to one. In addition, as autonomous agents become commonplace in public, they will likely need to interact with agents they have not previously encountered. These two facts provide a motivation to solve the Ad Hoc Teamwork problem, or the facilitation of collaboration between agents without any prior interaction. Many Ad Hoc Teamwork solutions do not consider the benefit provided by agents communicating. However, in practice communication channels may be common to many different agents (e.g. natural languages, various encoding methods, etc.), and it is not unreasonable to assume two agents that have not interacted before might already share a communication channel/protocol. In this thesis, we investigate a sub-task of Ad Hoc Teamwork, Communication in Ad Hoc Teamwork (CAT). In particular, we focus on a particular form of communication, queries, and introduce the Sequential Oneshot MultiAgent Limited Inquiry Communication in Ad Hoc Teamwork (SOMALI CAT) scenario. We first provide a theoretical analysis of when to ask queries in SOMALI CAT problems, followed by an empirical evaluation showing that asking at these times gives the best performance in practice. We then create principled algorithms for both optimizing the content of a query to ask the best possible question and interweaving querying into an agents plan. We provide experimental results to show that our methods provide a significant improvement over baselines.

Description

LCSH Subject Headings

Citation