engleski

Communication Service Optimization by Self-tranined and Distributed Agent Pools

The paper presents some simulated properties of the communication system supported by self-trained agents. The applied selection principle was major argument for the service properties in the designed agent environment. The work was motivated by redundancy, inherent in the variety of predicted abilities and properties of the network agents, particularly in distributed environments. Continuous adaptation was achieved by reinforcement Q-learning. Two options of possible environmental changes were studied: a significant change in SPA activity, and the increased traffic due to overloading of another node. Simulation results show rapid adaptation and significantly improved performance compared to sequential selection. As the consequence of the environmental change, the increase of the state time in the system has been detected, however always on the superior level than in the unlearned methods.

Reinforcement learning; Intelligent agents; Optimal routing

nije evidentirano