A resource distribution helps to improve the scalability and faulttolerance of many types of information systems, and can o.er the responsiveness needed by smart environments. When more than one information source is available, a load balancer distributes the user workload among the multiple sources. However, deciding the best server assignment for each user request is a di.cult task when the system contains heterogeneous elements and operates in a dynamic environment. We develop an automata-based approach to the load balancing problem that continuously adjusts the selection rate of servers based on observed information retrieval performance. We evaluate di.erent reinforcement methods to identify the most suitable one for this task, including the P-model Reward-Penalty-ϵ (RP-ϵ), and S-model Reward-Penalty-epsilon (SRP-ϵ). We develop a plug-in for Apache Tra.c Server and use a testbed to evaluate the performance of these methods compared to common load-balancing approaches: random, round robin, least connection, and greedy selection.