Although computer technology experienced significant growth during the 1970s, AI did not follow suit. In 1972, the UK’s primary science funding organisation, the Science Research Council, commissioned a report on AI’s progress that sparked considerable controversy within the field. Sir James Lighthill, a mathematician from Cambridge, authored a report outlining his conclusions and subsequently participated in a televised debate. In this debate, he faced off against three AI experts: Donald Michie and Richard Gregory from the University of Edinburgh, and John McCarthy from Stanford. Lighthill argued that AI researchers tended to oversimplify complex issues, claiming that AI represented poor science that consumed millions in funding without yielding satisfactory results after two decades of effort. The AI representatives in the debate attempted to defend their work against this public criticism, but with limited effectiveness.
AI research and funding went through an extended “wintering” period. However, during the 1970s, scientists achieved significant technological advancements that marked notable milestones for AI. We will examine several of these.
Shakey
At Stanford, computer scientists Charles Rosen, Nils Nilsson, and Peter Hart spearheaded the Shakey project. The robot, named Shakey due to its somewhat shaky movements as it navigated its surroundings, represented a groundbreaking advancement at the time. It integrated previously separate fields of AI research, including search, planning, natural language processing, and vision. In today’s terms, we would classify Shakey as a multimodal AI. However, it wasn’t particularly fast or intelligent; its processing power was housed elsewhere, as computers of that era were too bulky to be carried, necessitating a radio link for connectivity. Shakey operated solely within a specially designed environment filled with large, geometric shapes, such as cubes and pyramids. This setup echoed James Lighthill’s critique of oversimplifying problems, illustrating the challenges faced by researchers. Nonetheless, Shakey played a pivotal role in fostering several significant advancements in AI. The team devised the concept of a visibility graph, where each point (or node) represents a location and is linked by an edge if it is unobstructed. They also developed the A* search algorithm to determine subsequent actions and destinations. Additionally, they advanced the early field of computer vision with the Hough Transform, a method for feature extraction from camera images that identifies lines and ultimately shapes, such as circles. Shakey can be considered a precursor to contemporary autonomous robots and self-driving vehicles.
Genetic algorithms
In 1975, John Holland published a book about genetic algorithms. His concept was straightforward: if natural evolution could enhance life so that offspring generally had better adaptations to their environments than their parents, why couldn’t the same concepts be applied in computers to tackle problems? His genetic algorithms maintained a population of potential solutions to specific issues. Each solution consisted of a set of parameters that could be assessed using a fitness function to evaluate how effectively they addressed the issue. Solutions that achieved higher fitness scores were chosen preferentially as parents. These parent solutions would then produce a new generation of child solutions that inherited traits from their parents. Holland devised these digital genes using binary coding, where each parameter was represented as a binary number, allowing the resulting binary sequence of ones and zeros to be randomly combined with segments taken from each parent to create a child, incorporating a few random mutations for variety. Beginning with a random assortment of solutions, the genetic algorithm would progressively evolve increasingly effective solutions to the problem at hand.
Holland held positions as a professor of psychology, electrical engineering, and computer science at the University of Michigan. In 1977, he began experimenting with learning classifier systems. These systems were genetic algorithms capable of acquiring new rules in response to an evolving environment, with those rules applicable to tasks such as controlling robots or playing a game of draughts. A specialised bucket brigade algorithm identified which rules contributed to the success or failure of the classifier (rule) at any given moment. It allocated credit to the more successful ones, allowing them to produce more offspring while eliminating the less effective rules. Consequently, learning classifier systems could autonomously adapt to perform new tasks. This suggested something significant for artificial intelligence. Both genetic algorithms and learning classifier systems would grow into significant research domains that persist to this day.
Different chatbots
You may recall ELIZA, the very first chatbot developed in the 1960s. The exploration of chatbots progressed in the 1970s, when Kenneth Colby, a psychiatrist from Stanford, created PARRY. While ELIZA could imitate a doctor, PARRY was designed to act like a person suffering from paranoid schizophrenia. As part of a Turing Test in 1979, a panel of real psychiatrists was invited to interact with both actual patients and PARRY via teleprinters. They were tasked with distinguishing between the two, and they managed to do so correctly only 48 percent of the time. This indicated that PARRY was a more sophisticated chatbot than ELIZA, as it could mimic emotional reactions to specific prompts.
At MIT, Terry Winograd developed SHRDLU, named after the adjacent keys on a linotype machine keyboard. This AI functioned within a simplified blocks world and was capable of understanding and responding to inquiries about its environment, such as: “Is the red sphere on top of the green cube?” This block-world approach was the exact kind of simplification criticised by James Lighthill in his report a few years later.
Winograd later moved to Stanford, where he continued his research in AI and natural language understanding for many years. He also mentored two PhD students, Sergey Brin and Larry Page. Sergey Brin later recalled that Winograd had encouraged him to explore his idea of understanding the World Wide Web as a network of graphs. Following their research, Brin and Page founded Google in 1998. By that time, I was already utilising Google at Munich Re, conducting research for my books on insurance and reinsurance related to Workers’ Compensation, Employers’ Liability, and Occupational Diseases insurance systems around the globe.