EXTENDING GPGP FOR TEAMWORK OF A MULTI-AGENT SYSTEM

Abstract

In ·his appendix we present samples of the log files generated during a simulation rnn. Due to the distributed nature of multi-agent systems, each agent produces its own log file. To ease inspection, we implemented a utility program to 'merge' these individual log files into one combined log file where all events occurring at the same time at different agents are grouped together. Furthermore, the utility program only merges important events, tagged by the word "META" (thus the combined log file is to some extent abstracted). It should be noted that as we present the log file, agents may seem to change their beliefs (e.g. beliefs about a task, from achievable to unachievable and vice versa) in a frequent manner. This is mainly because of the truth maintenance system implemented in the Soar engine. For instance, when an agent starts up, unachievability conditions of tasks are not asserted yet, hence the agent believes all tasks to be achievable (tasks without unachievability conditions are always considered achievable). When unachievability conditions are asserted and before computing and propagating the expectations of each task, the agent believes all tasks to be unachievable. As more expectations are computed, more facts are asserted, and more tasks arc believed to be achievable.

The simulation run we are presenting has four properties: each team consists of four agents, both the synchronization mechanism and the monitoring framework are enabled (our approach), termination conditions indicates that at least two agents must complete their parts successfully, and degree of uncertainty D is used. Air team consists of al, a2, a3, and a4 (the order represents leadership replacement, i.e. a 1 is the leader but if it fails, a2 takes over the leadership and so on). Similarly, ground team consists of gl, g2, g3, and g4. To test leader failure recovery mechanism, we deliberately shut down a 1 in pulse 19. In the following sections we present portions of