This paper investigates the relative efficiency of two double-auction mechanisms for power exchanges, using agent-based modeling. Two standard pricing rules are considered and compared (i.e., "discriminatory" and "uniform") and computational experiments, characterized by different inelastic demand level, explore oligopolistic competitions on both quantity and price between learning sellers/producers. Two reinforcement learning algorithms are considered as well--"Marimon and McGrattan" and "Q-learning"--in an attempt to simulate different behavioral types. In particular, greedy sellers (optimizing their instantaneous rewards on a tick-by-tick basis) and inter-temporal optimizing sellers are simulated. Results are interpreted relative to game-theoretical solutions and performance metrics. Nash equilibria in pure strategies and sellers' joint profit maximization are employed to analyze the convergence behavior of the learning algorithms. Furthermore, the difference between payments to suppliers and total generation costs are estimated so as to measure the degree of market inefficiency. Results point out that collusive behaviors are penalized by the discriminatory auction mechanism in low demand scenarios, whereas in a high demand scenario the difference appears to be negligible.