Allocating Training Instances to Learning Agents that Improve Coordination for Team Formation

Agents can learn to improve their coordination with their teammates and increase team performance. We are interested in forming a team, i.e., selecting a subset of agents, that includes such learning agents. Before the team is formed, there are finite training instances that provide opportunities for the learning agents to improve. Agents learn at different rates, and hence, the allocation of training instances affects the performance of the team formed. We focus on allocating training instances to learning agent pairs, i.e., pairs that improve coordination with each other, with the goal of team formation. We formally define the learning agents team formation problem, and compare it with the multiarmed bandit problem. We consider learning agent pairs that improve linearly and geometrically, i.e., the marginal improvement decreases by a constant factor. We contribute algorithms that allocate the training instances, and compare against algorithms from the multi-armed bandit problem. In extensive simulations, we demonstrate that our algorithms perform similarly to the bandit algorithms in the linear case, and outperform them in the geometric case, thus illustrating the efficacy of our algorithms.