Achieving Human-Level Intelligence through Integrated Systems and Research: Introduction to This Special Issue
Cassimatis, Nicholas, Mueller, Erik T., Winston, Patrick Henry, AI Magazine
In many subfields of artificial intelligence during the past several decades, there has been substantial progress that has resulted in significant near-term advances in theory and applications. However, we believe that progress towards human-level artificial intelligence and the applications it enables requires a deeper and more comprehensive understanding that cannot be achieved by studying individual areas in isolation. Two reasons, both involving integration, support this belief. First, many problems that human-level AIs must solve involve subproblems currently addressed by different subfields, often using very different computational methods. A human-level AI must either integrate, for example, backtracking search, partially observable Markov decision processes (POMDPs), logic theorem proving algorithms, productions systems, and neural networks, or it must be based on new, heretofore undiscovered computational methods that exhibit all of the best features of these computational methods. Second, the best currently existing example of human-level intelligence is of course the human being. We believe that the history of AI (and the articles in this special issue) demonstrate that insights into the mechanisms underlying human cognition can inform research towards human-level AI. To explore a more integrated approach to human-level AI, we helped organize the 2004 AAAI Fall Symposium on Achieving Human-Level Artificial Intelligence through Integrated Systems and Research.
The symposium and the articles in this issue are motivated by four recent trends that we believe present an opportunity to reinvigorate the field's focus on understanding and developing human-level intelligence. First, many new subfields of artificial intelligence are explicitly directed towards integration. For example, many in the robotics community (see, for instance, the article by Craig Schlenoff, Jim Albus, Elena Messina, Tony Barbera, Raj Madhavan, and Stephen Balakirsky in this issue) are attempting to integrate the complexity of behavior enabled by reasoning and planning algorithms with the robustness and flexibility of reactive robotic architectures. Second, many new application domains require systems either to behave at the level of humans or to understand the complexities of human behavior. For example, the goal of increasing the realism of synthetic characters in training simulators (see the article by William Swartout, Jonathan Gratch, Randall Hill, Eduard Hovy, Stacy Marsella, Jeff Rickel, and David Traum) creates a need for artificial agents that behave at a human level of intelligence. Third, many funding agencies have recently renewed their interest in systems that go beyond the level of intelligence afforded by individual computational methods. Many of the programs resulting from the Defense Advance Research Project Agency's (DARPA) recent focus on cognitive systems, for example, aim to achieve a level of reasoning, deliberation, and learning normally associated with human intelligence. Finally, the ever increasing amount of computational power available to the field enables large-scale integration efforts that were heretofore impractical.
The articles in this issue report research that capitalizes on these trends by integrating multiple computational methods and by taking inspiration from recent discoveries about human cognition. As the work reported in these articles demonstrates, this approach has led both to deeper insights into intelligent systems and, consequently, to significant qualitative advances in the applications they enable.
Architectures That Combine the Insights of Multiple Subfields of AI
Subfields in AI are often based on different computational methods that are difficult to reconcile. However, since, for example, POMDPs, case-based reasoners, and logic-theorem provers each have strengths not shared by the others, human-level AI researchers must often find ways to integrate these different methods into a single intelligent system. …