Selective Perception Learning for Tasks Allocation

This paper presents a learning algorithm used to allocate tasks to agents in an uncertain real-time environment. In such environment, tasks have to be analyzed and allocated really fast for the multiagent system to be effective. To analyze those tasks, described by a lot of attributes, we have used a selective perception technique to enable agents to narrow down the description of each task by choosing the attributes that it should be considering in each situation. By doing so, we have obtained a drastic reduction of the number of possible states. We have used this algorithm at two different levels for the problem of choosing the best fire to extinguish for each firefighter agent in the RoboCupRescue simulation environment. First, a center agent is using the algorithm to allocate a zone on fire for each firefighter agent. Then, those agents are choosing the best fire to extinguish in this zone. Our results show a good improvement in the agents capability to extinguish fires, as the agents become better at distinguishing the world states.