Computational tradeoffs under bounded resources

Eric Horvitz, Shlomo Zilberstein
2001 Artificial Intelligence  
Over the nearly fifty years of research in Artificial Intelligence, investigators have continued to highlight the computational hardness of implementing core competencies associated with intelligence. Key pillars of AI, including search, constraint propagation, belief updating, learning, decision making, and the associated real-world challenges of planning, perception, natural language understanding, speech recognition, and automated conversation continue to make salient the omnipresent wall of
more » ... computational hardness. Early pioneers in AI research, including Allen Newell and Herbert Simon, established a long tradition of battling obvious intractabilities by resorting to approximations that relied on heuristic procedures-informal policies that appeared to perform acceptably on subsets of real-world problems. Bounded rationality was conceived and popularized in the context of sample applications that relied on such heuristic procedures to struggle through overwhelming complexity. In the mid-1980s, several researchers began to pursue a line of research aimed at better understanding and formalizing tradeoffs under bounded representational and computational resources. During this time, a palpable shift in perspective occurred with regard to tackling resource limitations. Rather than viewing scarce resources as an unfortunate impediment, foiling at every turn attempts to perform automated problem solving on realistic challenges, investigators began to consider tradeoffs under scarce resources as a rich arena for focused AI research. Passionate researchers suggested that elusive principles of intelligence might actually be founded in developing a deeper understanding of how systems might grapple, in an implicit or explicit resource aware manner, with scarce, varying, or uncertain time and memory resources. Beyond computational resources, the interaction of limited resources and constraints associated with fixed problem-solving architectures were explored. Older, informal notions of bounded rationality soon gave way to richer, more comprehensive approaches to rational computational and real-world actions that incorporate considerations of resource costs and constraints.
doi:10.1016/s0004-3702(01)00051-0 fatcat:jberehb5sbbfpacgmm6daqtlmy