Vincent Conitzer and Tuomas Sandholm
We present BL-WoLF, a framework for learnability in repeated zero-sum games where the cost of learning is measured by the losses the learning agent accrues (rather than the number of rounds). The game is adversarially chosen from some family that the learner knows. The opponent knows the game and the learner’s learning strategy. The learner tries to either not accrue losses, or to quickly learn about the game so as to avoid future losses (this is consistent with the Win or Learn Fast (WoLF) principle; BL stands for "bounded loss"). Our framework allows for both probabilistic and approximate learning. The resultant notion of BL-WoLF-learnability can be applied to any class of games, and allows us to measure the inherent disadvantage to a player that does not know which game in the class it is in. We present guaranteed BL-WoLF-learnability results for families of games with deterministic payoffs and families of games with stochastic payoffs. We demonstrate that these families are guaranteed approximately BL-WoLF-learnable with lower cost. We then demonstrate families of games (both stochastic and deterministic) that are not guaranteed BL-WoLF-learnable. We show that those families, nevertheless, are BL-WoLFlearnable. To prove these results, we use a key lemma which we derive.