An auction-based approach to spectrum allocation using multi-agent reinforcement learning

We present an auction-based approach to spectrum management in a multi-operator context. Service providers compete for customers in real-time through live auctions. To automate the bidding process we implement a multi-agent reinforcement learning solution. We study the effect of real-time competition between service providers by considering the cases where there is a single provider and multiple providers. Furthermore, we demonstrate how users of varying types, based on application-type and willingness to pay, can be accommodated. We utilize a low-complexity bid-proportional allocation mechanism which ensures fairness. Our simulation results show that when there is a single provider, revenue can be maximized by artificially limiting supply and creating contention. However, when there are multiple providers from which the customers can dynamically choose, there is no longer an incentive to restrict supply due to the direct competition between service providers.