With access to electronic digital programmable computers in the mid-1950s, AI researchers began to focus on symbol manipulation (i.e., the manipulation of mathematical expressions, as is found in algebra) to emulate human intelligence. Three institutions led the charge: Carnegie Mellon University, Stanford University, and the Massachusetts Institute of Technology (MIT). Each university had its own style of research, which the American philosopher John Haugeland (1945–2010) named “good old-fashioned AI” or “GOFAI.”
From the 1960s through the 1970s, symbolic approaches achieved success at simulating high-level thinking in specific application programs. For example, in 1963, Danny Bobrow’s technical report from MIT’s AI group proved that computers could understand natural language well enough to solve algebra word problems correctly. The success of symbolic approaches added credence to the belief that symbolic approaches eventually would succeed in creating a machine with artificial general intelligence, also known as “strong AI,” equivalent to a human mind’s intelligence.
By the 1980s, however, symbolic approaches had run their course and fallen short of the goal of artificial general intelligence. Many AI researchers felt symbolic approaches never would emulate the processes of human cognition, such as perception, learning, and pattern recognition. The next step was a small retreat, and a new era of AI research termed “subsymbolic” emerged. Instead of attempting general AI, researchers turned their attention to solving smaller specific problems. For example researchers such as Australian computer scientist and former MIT Panasonic Professor of Robotics Rodney Brooks rejected symbolic AI. Instead he focused on solving engineering problems related to enabling robots to move.
In the 1990s, concurrent with subsymbolic approaches, AI researchers began to incorporate statistical approaches, again addressing specific problems. Statistical methodologies involve advanced mathematics and are truly scientific in that they are both measurable and verifiable. Statistical approaches proved to be a highly successful AI methodology. The advanced mathematics that underpin statistical AI enabled collaboration with more established fields, including mathematics, economics, and operations research. Computer scientists Stuart Russell and Peter Norvig describe this movement as the victory of the “neats” over the “scruffies,” two major opposing schools of AI research. Neats assert that AI solutions should be elegant, clear, and provable. Scruffies, on the other hand, assert that intelligence is too complicated to adhere to neat methodology.
From the 1990s to the present, despite the arguments between neats, scruffies, and other AI schools, some of AI’s greatest successes have been the result of combining approaches, which has resulted in what is known as the “intelligent agent.” The intelligent agent is a system that interacts with its environment and takes calculated actions (i.e., based on their success probability) to achieve its goal. The intelligent agent can be a simple system, such as a thermostat, or a complex system, similar conceptually to a human being. Intelligent agents also can be combined to form multiagent systems, similar conceptually to a large corporation, with a hierarchical control system to bridge lower-level subsymbolic AI systems to higher-level symbolic AI systems.
The intelligent-agent approach, including integration of intelligent agents to form a hierarchy of multiagents, places no restriction on the AI methodology employed to achieve the goal. Rather than arguing philosophy, the emphasis is on achieving results. The key to achieving the greatest results has proven to be integrating approaches, much like a symphonic orchestra integrates a variety of instruments to perform a symphony.
In the last seventy years, the approach to achieving AI has been more like that of a machine gun firing broadly in the direction of the target than a well-aimed rifle shot. In fits of starts and stops, numerous schools of AI research have pushed the technology forward. Starting with the loftiest goals of emulating a human mind, retreating to solving specific well-defined problems, and now again aiming toward artificial general intelligence, AI research is a near-perfect example of all human technology development, exemplifying trial-and-error learning, interrupted with spurts of genius.
Although AI has come a long way in the last seventy years and has been able to equal and exceed human intelligence in specific areas, such as playing chess, it still falls short of general human intelligence or strong AI. There are two significant problems associated with strong AI. First, we need a machine with processing power equal to that of a human brain. Second, we need programs that allow such a machine to emulate a human brain.