Decision-making in uncertainty and coordination are at the heart of multiagentsystems. In this kind of systems, agents have to be able to perceive their environmentand take decisions while considering the other agents. When the environment is partiallyobservable, agents have to be able to manage this uncertainty in order to take themost enlightened decisions they can based on the incomplete information they haveacquired. Moreover, in the context of cooperative multiagent environments, agentshave to coordinate their actions in order to accomplish complex tasks requiring morethen one agent.In this thesis, we consider complex cooperative multiagent environments (dynamic,uncertain and real-time). In this kind of environments, we propose an approach ofdecision-making in uncertainty that enable the agents to flexibly coordinate themselves.More precisely, we present an online algorithm for partially observable Markov decisionprocesses (POMDPs).Furthermore, in such complex environments, agent's tasks can also become quitecomplex. In this context, it could be complicated for the agents to determine therequired number of resources to accomplish each task. To address this problem, wepropose a learning algorithm to learn the number of resources necessary to accomplisha task based on the characteristics of this task. In a similar manner, we propose ascheduling approach enabling the agents to schedule their tasks in order to maximizethe number of tasks that could be accomplish in a limited time.All these approaches have been developed to enable the agents to efficiently coordinateall their complex tasks in a partially observable, dynamic and uncertain multiagentenvironment. All these approaches have demonstrated their effectiveness in tests donein the RoboCupRescue simulation environment.
