Decentralized Control of Large-Scale Networks as a Game with Local Interactions: Cross-layer TCP/IP Optimization

Developing optimized distributed protocols for large-scale networks is a challenging problem due to scalability and stability concerns. Scalability concerns can be naturally addressed by interpreting distributed protocols as a non-cooperative game of local protocol components attempting to maximize their individual utilities. One of the difficulties in implementing this approach is developing adaptive algorithms capable of learning of the expected utilities and adjusting the corresponding control actions for the purpose of approaching the solution to the corresponding game, and thus optimization of the global system performance. It is known that the best response by each component to its expected utility may result in unstable behavior and deterioration of the overall performance. On an example of cross-layer optimization of a TCP/IP network, this paper discusses the possibility of avoiding these undesirable effects by allowing the control actions occasionally deviate from their best response values. Using simulations, the paper suggests that (a) sufficient level of randomness in route selection improves the network performance by eliminating the route flapping instability, (b) the optimal level of randomness keeps the network within the stability region in close proximity to the border of this region, and (c) it may be possible to optimize the network performance by adjusting the level of randomness.