No regrets about no-regret

摘要

No-regret is described as one framework that game theorists and computer scientists have converged upon for designing and evaluating multi-agent learning algorithms. However, Shoham, Powers, and Grenager also point out that the framework has serious deficiencies, such as behaving sub-optimally against certain reactive opponents. But all is not lost. With some simple modifications, regret-minimizing algorithms can perform in many of the ways we wish multi-agent learning algorithms to perform, providing safety and adaptability against reactive opponents. We argue that the research community should have no regrets about no-regret methods.