Date of Original Version
This is the author’s version of a work that was accepted for publication. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version is available at http://dx.doi.org/10.1016/j.artint.2013.12.002
Abstract or Description
Previous approaches to select agents to form a team rely on single-agent capabilities, and team performance is treated as a sum of such known capabilities. Motivated by complex team formation situations, we address the problem where both single-agent capabilities may not be known upfront, e.g., as in ad hoc teams, and where team performance goes beyond single-agent capabilities and depends on the specific synergyamong agents. We formally introduce a novel weighted synergy graph model to capture new interactions among agents. Agents are represented as vertices in the graph, and their capabilities are represented as Normally-distributed variables. The edges of the weighted graph represent how well the agents work together, i.e., their synergy in a team. We contribute a learning algorithm that learns the weighted synergy graph using observations of performance of teams of only two and three agents. Further, we contribute two team formation algorithms, one that finds the optimal team in exponential time, and one that approximates the optimal team in polynomial time. We extensively evaluate our learning algorithm, and demonstrate the expressiveness of the weighted synergy graph in a variety of problems. We show our approach in a rich ad hoc team formation problem capturing a rescue domain, namely the RoboCup Rescue domain, where simulated robots rescue civilians and put out fires in a simulated urban disaster. We show that the weighted synergy graph outperforms a competing algorithm, thus illustrating the efficacy of our model and algorithms.
Artificial Intelligence, 208, 41-65.