Exact Learning Algorithms, Betting Games, and Circuit Lower Bounds

John M. Hitchcock and Ryan C. Harkins

Abstract:
This paper extends and improves work of Fortnow and Klivans [2009], who showed that if a circuit class C has an efficient learning algorithm in Angluin's model of exact learning via equivalence and membership queries [Angluin, 1988], then we have the lower bound EXPNPC. We use entirely different techniques involving betting games [Buhrman et al., 2001] to remove the NP oracle and improve the lower bound to EXP ⊈ C. This shows that it is even more difficult to design a learning algorithm for C than the results of Fortnow and Klivans indicated. We also investigate the connection between betting games and natural proofs, and as a corollary the existence of strong pseudorandom generators.

Full Version:

Preliminary Version: