BELMONT, MASSACHUSETTS – Symbolic artificial intelligence (AI) in the 1970s and 1980s focused on imparting to machines knowledge about narrowly defined virtual worlds and about specific domain expertise. In parallel, new statistical approaches to learning from data also emerged, demonstrating their practical value.

Microworlds

At the Massachusetts Institute of Technology, cognitive and computer scientist Marvin Minsky and a few of his students used funds from United States government agencies to work on a handful of narrow problems, developing AI programs that could function in virtual spaces they called “microworlds.” Computer scientist Terry Winograd’s SHRDLU (1968) presents a good example: A natural language understanding computer program with which the user can converse with a computer in ordinary English, instruct it to move or place blocks in different places and positions in a virtual environment, and ask it questions about them.

Moving from microworlds and narrowly defined problems to larger real-world environments and complex problems, however, proved to be very difficult. The assumption that the mere availability of faster hardware and larger memories would be enough for ‘scaling up’ turned out to be too optimistic. 

A 1973 report on the state of artificial intelligence (AI) in the United Kingdom described the failure to apply research results to real-world problems and convinced the UK government to end support for AI research in all but two universities. Similar disappointment from the failure of such research to deliver on overhyped promises led the Defense Advanced Research Projects Agency (DARPA) in the US to cut its AI funding.

The dismay that funding agencies expressed over the state of AI was shared by a few prominent AI researchers like Winograd, who abandoned symbolic AI in favor of pursuing work d