Gaming and Learning Approaches for Multi-User Computation Offloading

We consider both offline and online computational offloading of tasks from multiple users to a cloud or nearby cloud at the edge. We model the offline problem as an -player finite game where each user has access to information from other users, and we use an optimization approach to find a mixed-strategy Nash equilibrium solution. We also consider a practical online version wherein tasks arrive over time and a user does not require information from other users. We suggest a solution to this online problem by adopting a payoff-based reinforcement learning algorithm, which converges to a pure-strategy solution. Through simulation, we observe that the trends of the Nash equilibrium obtained from the offline technique and the pure-strategy point obtained from the online solution are similar. While the offline algorithm obtains a better solution on average, the online algorithm is much faster, particularly for larger systems.