Provably bounded optimal agents

摘要: Since its inception, arti cial intelligence has relied upon a theoretical foundation centred around perfect rationality as the desired property of intelligent systems. We argue, others have done, that this is inadequate because it imposes fundamentally unsatis able requirements. As result, there arisen wide gap between theory and practice in AI, hindering progress eld. propose instead called bounded optimality. Roughly speaking, an agent bounded-optimal if program solution to constrained optimization problem presented by architecture task environment. show how construct agents with for simple class machine architectures broad real-time environments. illustrate these results using model automated mail sorting facility. also de ne weaker property, asymptotic optimality (ABO), generalizes notion classical complexity theory. then universal ABO programs, i.e., programs are no matter what constraints applied. Universal can be used building blocks more complex conclude discussion prospects basis relate similar trends philosophy, economics, game