Abstract
In this paper we examine combined fully distributed payoff and strategy learning (CODIPAS) in a queue-aware access game over a graph. The classical strategic learning analysis relies on vanishing or small learning rate and uses stochastic approximation tool to derive steady states and invariant sets of the underlying learning process. Here, the stochastic approximation framework does not apply due to non-vanishing learning rate. We propose a direct proof of convergence of the process. Interestingly, the convergence time to one of the global optima is almost surely finite and we explicitly characterize the convergence time. We show that pursuit-based CODIPAS learning is much faster than the classical learning algorithms in games. We extend the methodology to coalitional learning and proves a very fast formation of coalitions for queue-aware access games where the action space is dynamically changing depending on the location of the user over a graph.
Original language | English (US) |
---|---|
Title of host publication | IFAC Proceedings Volumes |
Publisher | Elsevier BV |
Pages | 7055-7060 |
Number of pages | 6 |
ISBN (Print) | 9783902823625 |
DOIs | |
State | Published - Oct 8 2014 |