Distributed Problem Solving(DPS), as a part of Distributed Artificial Intelligence(DAI), deals with the interactions of agents attempting to solve problems together. The agent which is a problem solver has only partial information for global environment, and cannot solve a whole problem by itself. Many DPS systems have been developed using the concept from the distributed processing architecture and the AI rule representation. But a few systems take formalism-based approach. As the DPS maintains discrete event system property, we propose a DPS framework which is modeled by Discrete Event System Specification(DEVS) formalism and implemented in DEVSim++ which is a simulation tool dereloped in oure laboratory. The models of the framework classify both behavior and structure according to the character of the agent. Therefore, the operational meaning of an agent model can be exchanged by some representations. Our DPS system constructed flexibly our in modular, hierarchical manner and its model can be reused according to their applications. Using framework, we tested three manufacturing systems with respect to distributed control, failure/repair operation and a degree of redundancy. In the DPS, the knowledge source of the agent is an important element. The capability of problem-solving is determined by the knowledge source. Recently, neural network has been employed to the rule representation of knowledge sources. Back-propagation neural network is a kind of neural network. A significant problem is the training efficiency for acquiring knowledge source with the back-propagation neural network. In order to effectively acquire knowledge source, we suggest two training methods(on-line and window control) of the neural network. These methods are applied to evolutionary programming technique. The objective function and perturbation factor of evolutionary programming algorithm is modified by the parameters of the network. During two benchmark and one modulus problem tests, our methods were superior to otheradaptive training methods: quasi-newton, steepest-descent and weight-average methods. Specially, as the complexity of structure increases, our methods are much efficient. Merging these learning algorithms to our DPS frame work is remained further work.