Using Dynamic Neural Network to Model Team Performance for Coordination Algorithm Configuration and Reconfiguration of Large Multi-Agent Teams

Coordination of large numbers of agents for performing complex tasks in complex domains is a rapidly progressing area of research. Because of the high complexity of the problem, approximate and heuristic algorithms are typically used for key coordination tasks. Such algorithms usually require tuning of algorithm parameters to get the best performance in particular circumstances. Manually tuning of parameters is sometime difficult. In this paper, we introduce a new concept of dynamic features for a neural network, called dynamic networks, to model the way a coordination algorithm will work under particular circumstances. Genetic algorithms are used to train the networks from an abstract simulation, TeamSim. At the end, the model is used to rapidly determine an appropriate configuration of the algorithm for a particular domain. Users specify required tradeoffs in algorithm performance and use the neural network to find the best configuration for those tradeoffs. Algorithm reconfiguration can even be performed online to improve the performance of an executing team as situation changes. We present preliminary results showing the approach promisingly facilitating users to configure and control a large team executing sophisticated teamwork algorithms. INTRODUCTION AND PREVIOUS WORK Sophisticated, complex coordination allows large groups of agents to perform complex tasks in domains such as space (Kortenkamp et al., 2000) and the military (Glade, 2000). Cooperation between heterogeneous robot teams is very complex when the tasks cannot be completely divided up and assigned a priori and robots must work together, dynamically interacting to achieve common goals, for example see work of Parker et al (2001). In a large team (Scerri et al., 2004), the teamwork algorithms include with several coordination algorithms, such as algorithms for task allocation, communication, and planning. The relationships between these algorithms to the behaviors of team are highly non-linear, stochastic, and extremely complex. Knowing very well only in a part of the system is not likely to understand how the team can be controlled and get the best performance out of them. Therefore, it is critical for human operators to have better understanding of overall teamwork algorithms. However, this broadly understanding can not easily obtain by running a few of the real missions. In multi-robot systems, it is usually impossible to have lots of physical experiments, therefore simulation is used to gain understanding about the developing system ((Dixon et al., 1999), (Fu et al., 2003)). On the other hand, using simulation is difficult when the