A discussion of a fusion problem in multi-agent systems for time critical decision making is presented. The focus is on the problem of distributed learning for classification into several hypotheses of observations representing states of an uncertain environment. Special attention is devoted to reinforcement learning in a homogeneous non-communicating multi-agent system for time critical decision making. A system in which an agent network processes observational data and outputs beliefs to a fusion center module is considered. Belief theory serves as the analytic framework for computing these beliefs and composing them over time and over the set of agents. The agents are modeled using evidential neural networks, whose weights reflect the state of learning of the agents. Training of the network is guided by reinforcements received from the environment as decisions are made. Two different sequential decision making mechanisms are discussed: the first one is based on a “pignistic ratio test” and the second one is based on “the value of information criterion,” providing for learning utilities. Results are shown for the test case of recognition of naval vessels from FLIR image data.