The thesis's subject is a communicating system which uses a general load-balancing between processes, or computers. A lot of applications reach a level that we would like to implement/operate an architecture or a network from it's instances to apply better scalability and optimize its performance. In this case one of the biggest challenges is that the nodes, and the links of the network could fail at any time, and could also resurrect easily. In favor of the most reachable adaptivity I am using hierarchical multi-agent reinforcement learning, and also neural-networks for estimating usefulness.
Compared to the original routing protocols, the final aim is to create a network which transfer capability is at least as good as theirs. In addition in my version I want to implement it with much simpler, and solid code-base and also to take the advantages of the network's all unique properties by the learning. The framework would be a system which implements a network using socket communication, it's code-base is as simple as it is possible, and it can pass along any message from „A” to „B” as fast as it possible. The network performs all these duties with the constraint that it can only send a package to an other node if it is allowed, and also can achieve load-balancing using the statistics from the other nodes (it can also helps the fast package forwarding).