Cognitive Networks: Towards Self-Aware Networks

4.5. Self-Management Advances in Specific Problem Domains

4.5.1. Self-Optimization

In ^[] utility functions are used for the self-optimization of data center servers. The purpose of the utility function is to provide to data service requests the best resource allocation among a group of data servers. The utility computation itself is performed in two levels. At the local level, the autonomic application manager determines the utility of servicing a specific selection of requests based on its fixed resources and demand requirements. At the global level, allocation of resources to application managers is determined by a global resource arbiter.

Collaborative reinforcement learning is used in a k-component model to generate self-adaptive distributed systems ^[]. The k-component is a modeling framework for creating distributed components and specifying their interfaces. Collaborative reinforcement learning is a modification of the reinforcement learning approach, which allows the components to not only learn from environmental feedbacks but also from the experiences of neighboring components. Global optimization problems are thus tackled in a distributed manner by first having the individual component perform its local reasoning using reinforcement learning and then advertise the results to the neighbors. Based on the results, the neighbors may initiate another round of distributed optimization process. The dynamic control of behavior based on learning (DCBL) middleware ^[] uses reinforcement ...