Team-partitioned, opaque-transition reinforcement learning (TPOT-RL) is a distributed reinforcement learning technique that allows a team of independent agents to learn a collaborative task. TPOT-RL was first successfully applied to simulated robotic soccer. This paper demonstrates that TPOT-RL is general enough to apply to a completely different domain, namely network packet routing. Empirical results in an abstract network routing simulator indicate that agents situated at individual nodes can learn to efficiently route packets through a network that exhibits changing traffic patterns, based on locally observable sensations.

PDF, PS, HTML

In Proceedings of the Seventeenth International Conference on Machine Learning, 935-942, 2000.

@InProceedings{ICML2000, title={TPOT-RL Applied to Network Routing}, author={Peter Stone}, booktitle={Proceedings of the Seventeenth International Conference on Machine Learning}, pages={935-942}, url="http://nn.cs.utexas.edu/?ICML2000", year={2000} }