Fifth Generation Computer Systems project: Japan bets a decade on logic machines

The event and the date

In October 1981, Japan's Ministry of International Trade and Industry (MITI) used a four-day international conference in Tokyo, held from 19 to 22 October, to present a plan for what it called Fifth Generation Computer Systems, according to ICOT's 1984 status report by Kazukiyo Kawanobe. The conference proceedings, published by the Japan Information Processing Development Center, survive in collections such as the Computer History Museum's catalog. Six months later the plan became an institution. The Institute for New Generation Computer Technology, known as ICOT, was formally launched on 14 April 1982, as Edward Feigenbaum and Pamela McCorduck recorded at the time, and its research center began work that June, with MITI paying the entire cost under contract, according to Takashi Kurozumi's ten-year overview of the project.

Chronicle dates the event to 1982 because that is when the budget, the institute and the ten-year clock started. The 1981 conference was the announcement, and it is the moment the rest of the world reacted to, so both dates matter, but the project itself is a 1982 event.

The narrow claim is simple to state. A government ministry committed about 54 billion yen, over what became 11 years, to building computers whose basic operation would be logical inference rather than arithmetic, organized around one logic programming language and aimed at what its planners called knowledge information processing, as the IPSJ Computer Museum summarizes. The broad claims are harder. The project is remembered as a cautionary tale of AI ambition, as the catalyst for rival programs in Britain, Europe and the United States, and as a quiet success in parallel logic programming. Each reading has evidence behind it, and this report keeps them apart.

A committee's case for a new kind of computer

The idea predates ICOT by three years. In 1979, at MITI's suggestion, the Japan Information Processing Development Center set up a committee to define a fifth generation of computers. Over well more than one hundred meetings it weighed inference machines for knowledge processing, database and knowledge-base machines, high-performance workstations, distributed systems and scientific supercomputers, as Kurozumi's overview recounts. The same account explains the policy mood. Japanese computer makers had reached the level of the best foreign machines, and the national project was meant to shift from catching up toward a risky, original contribution to world science.

The committee's answer was specific. The new computer would rest on parallel, non-von Neumann processing and on inference over knowledge bases, with a logic programming language as the interface between hardware and software. The stated goal was a prototype of about a thousand element processors running at 100 million to 1 billion logical inferences per second (LIPS), reached over ten years split into a three-year initial stage, a four-year intermediate stage and a three-year final stage ( Kurozumi). For scale, Nils Nilsson notes in his history of the field that computers of the early 1980s managed around 100,000 LIPS, so the planners were reaching for a thousandfold gain or more.

Kazuhiro Fuchi, who directed the ICOT research center, presented the design as a hypothesis to be tested. In his 1984 restatement of the project's philosophy, he described the target as a predicate logic machine whose kernel language would play the role a machine language plays on a conventional computer, with knowledge representation languages, and eventually natural language, layered above it. He insisted that this was not an artificial intelligence or expert-system project, as some outsiders assumed, while conceding that its premise was a forecast that AI applications would become the mainstream of computing. Programs such as DENDRAL had shown that encoded knowledge could do expert work, and the committee bet that hardware built for inference could carry far more of it.

How an inference machine was meant to work

Logic programming writes a program as facts and rules. One fact might say that Hanako is a parent of Taro, another that Taro is a parent of Jiro, and a rule might say that X is a grandparent of Z if X is a parent of Y and Y is a parent of Z. Asked whose grandparent Hanako is, the system matches the question against the rule, then matches each condition against the facts, a step called resolution. Each such step counts as one logical inference, which is why the project measured speed in LIPS rather than arithmetic operations. Prolog, the language that made this practical in the 1970s, tries alternatives one at a time and backtracks when one fails. That is easy to run on one processor and awkward to spread across hundreds.

ICOT's first stage built tools for the sequential case. Its Personal Sequential Inference machine, PSI, ran a Prolog-derived kernel language called KL0 at about 35,000 LIPS, and ICOT wrote an operating system for it, SIMPOS, in an extended logic language called ESP: more than 100,000 lines produced by about twenty people in roughly two years, according to Shunichi Uchida's summary of the parallel inference machine. Uchida reports that the experience convinced the team that even an operating system could be written productively in a logic language, which answered one of the early criticisms Fuchi had acknowledged.

Parallelism required a new language. Prolog's backtracking search fits poorly with many processors working at once, which is why, as Nilsson notes, the project turned to Guarded Horn Clauses (GHC), a concurrent logic language. Kurozumi credits GHC to discussions with researchers working on two earlier concurrent logic languages, Parlog and Ehud Shapiro's Concurrent Prolog ( overview). In GHC and its practical successor, the kernel language KL1, each clause carries a guard. A process waits until its inputs are bound enough to decide which clause applies, commits to that clause, and never backtracks. Synchronization falls out of the data itself, and Uchida describes KL1 as combining automatic memory management with dataflow synchronization ( Uchida). The hard part moved into the operating system, PIMOS, which had to split irregular symbolic workloads and balance them across processors.

The hardware followed in steps. A Multi-PSI system of 64 processors on an eight-by-eight mesh, completed in spring 1988, ran KL1 at about 150,000 LIPS per processor and 5 to 10 million LIPS overall ( Uchida). The final prototype was a family of five Parallel Inference Machine (PIM) models with different architectures and about a thousand processors in all. The largest, PIM/p, linked up to 512 processing elements in clusters of eight over a hypercube network, and PIM/m linked 256 over a two-dimensional mesh, according to ICOT's archived specifications.

Alarm abroad

The October 1981 announcement landed in a West already anxious about Japanese manufacturing, and it came with an invitation to take part. Britain's reaction is the best documented. A committee chaired by John Alvey was set up in March 1982, and its report stated that the catalyst for the committee was the unveiling of Japan's program and Japan's invitation to discuss participation. A British delegation had attended the Tokyo conference and reported to a London meeting of industrialists, who saw the scale and cohesion of the Japanese effort as a major competitive threat, as the Alvey Report records. By August 1983 the government had approved a five-year program of 350 million pounds, 200 million of it public money, according to the Christian Science Monitor. The same article noted that the original Japanese invitation had been met mainly with scorn by Western researchers, who expected Japan to take more than it gave.

In the United States the reaction was louder and more diffuse. Feigenbaum and McCorduck's 1983 book, The Fifth Generation, framed the project as a national challenge. In an excerpt printed in Creative Computing, they wrote that MITI planned about 450 million dollars of seed money over the decade, with industry expected to match or double it, and called the plan an exquisite piece of economic strategy. (ICOT's own accounts later state that MITI paid the whole cost.) Nilsson traces the alarm to American industry, where a February 1982 meeting of executives led to the Microelectronics and Computer Technology Corporation, a research consortium in Austin, and records that the European Community launched its ESPRIT program in 1983 partly as a response ( Nilsson). He describes DARPA's Strategic Computing program as partly a response to the Japanese project and partly the product of DARPA's own agenda, and cites the historian Alex Roland's judgment that Congress was more exercised by the Fifth Generation than either the administration or the computer community ( Nilsson on Strategic Computing).

These links deserve to be stated at their documented strength. For Alvey, the Japanese program is the stated catalyst in the founding report itself. For the American consortium and for ESPRIT, it was one cause among several, alongside older worries about Japanese strength in semiconductors and consumer electronics. For Strategic Computing, it helped sell the program to Congress more than it shaped DARPA's plan.

Visiting researchers saw another motive. Ehud Shapiro, whose Concurrent Prolog fed ICOT's language work, wrote a trip report for Communications of the ACM presenting the project as a revolutionary ten-year plan, and reporting that it intended to answer the charge that Japan used foreign knowledge without contributing its own by publishing its results for the international community.

What ICOT built

The money arrived roughly as planned. Kurozumi gives about 8 billion yen for the three-year initial stage, about 22 billion for the four-year intermediate stage, about 21 billion for 1989 to 1991 and an estimated 3.6 billion for 1992, when the final stage was stretched to four years to evaluate the whole system, for a total of about 54 billion yen over 11 years ( overview). Nilsson converts the total to about 380 million dollars at the 1990 exchange rate ( Nilsson). Researchers were seconded from national laboratories and computer makers, limited to age 35 and rotated after three to four years; the research center grew from 40 people in 1982 to about 100, and 184 researchers passed through it.

The program narrowed as it went. At the end of the initial stage, machine translation, speech, and picture and image processing were dropped from the project because computer vendors were already active in them, and work on a large electronic dictionary moved to a separate organization. Each stage had one standard software environment, Prolog first, then ESP, then KL1, so that tools could be shared ( Kurozumi). Applications written in KL1 for the final machines included chip design tools, protein sequence analysis, legal reasoning, a parallel theorem prover and programs for the game of Go ( overview).

The project was also unusually open for a national industrial program. By January 1992 ICOT had issued more than 700 technical reports and 1,100 technical memos, sent its journal to more than 600 locations in 35 countries and hosted 74 invited researchers from 12 countries. Its 1984 and 1988 conferences drew about 1,100 and 1,600 participants ( Kurozumi).

Two verdicts on the same machines

When the final conference opened in Tokyo on 1 June 1992, as the ICOT archive's conference listing dates it, two incompatible verdicts appeared almost at once. Computer Business Review reported on 2 June 1992 that MITI was closing the project after ten years and about 400 million dollars, that the ministry had admitted falling short of computers that reason the way people do, and that most participating companies had built only one-off prototypes rather than products. ICOT's own account, in a 1994 letter to the Australian National University, described the project as concluded in 1992 after achieving its initial goal.

Both were measuring something real. The official final evaluation of March 1993, written by an evaluation working group of the committee that supervised the project's progress, found that the PIM exceeded 150 million LIPS against an original goal of 100 million. Feigenbaum and Shrobe's 1993 survey likewise put measured performance in the planned range and called KL1 an excellent system for logic programming on parallel machines ( Feigenbaum and Shrobe). Uchida adds the necessary caveat. Writing in March 1992, as the 256-processor PIM/m was just starting to run benchmarks, he expected it to exceed 100 million LIPS at peak and warned that real applications could run much slower depending on their structure ( Uchida).

Measured against the vision the world heard in 1981, the outcome looks different. Feigenbaum and McCorduck had relayed a promise of machines that would converse in natural language and understand speech and pictures ( excerpt). The same final evaluation conceded that handling speech and images as knowledge rather than numerical data required accumulating and structuring vast background knowledge. From the intermediate stage onward the work concentrated on natural language and knowledge representation, and no concrete interface system was built ( final evaluation). Feigenbaum and Shrobe reported that applications existed only as demonstrations, and that the lack of applications customers found effective, together with sole reliance on a language outside the mainstream, disenchanted the Japanese computer makers. The project met its engineering targets and missed the promise that had made it famous.

Why the bet missed

Three explanations recur in the record. The first is hardware economics. Nilsson observes that commercial workstations, and even personal computers, improved fast enough to overtake the special-purpose PIMs, and that exploiting parallelism proved hard outside problems naturally suited to it ( Nilsson). ICOT's own projections show the squeeze. Kurozumi projected that a single board would reach about 20 million LIPS, a hundred times a PIM, so that on ICOT's own trend lines the speed of the whole prototype would before long fit on a handful of boards ( overview). A dedicated architecture has to stay ahead of general-purpose processors to justify its cost, and in the 1980s general-purpose processors moved very fast.

The second is the language bet. Fuchi had called predicate logic a working hypothesis, and it produced elegant systems, but Feigenbaum and Shrobe single out the sole use of a language outside the mainstream, alongside the missing applications, as the reason the Japanese computer makers lost interest ( Feigenbaum and Shrobe). The third is the knowledge problem the final evaluation itself identified. Inference speed was never the binding constraint on understanding speech or pictures; the missing ingredient was knowledge, and more LIPS could not supply it. Knowledge-based programs from DENDRAL onward had already met a related wall, the slow and costly work of encoding expertise by hand.

The project's relation to the second AI winter needs similar care. Nilsson places that winter in the mid to late 1980s, as sponsors' inflated expectations met systems that could not match them ( Nilsson on the AI winter). The Fifth Generation was among the most conspicuous of those expectations, and the alarm it caused abroad helped inflate the boom. But the project ran to 1992 on a steady budget, and its official end came after the commercial downturn in AI had already arrived. It is best read as part of the same cycle of promise and disappointment as the first AI winter, and as a symbol of the second, rather than as its single cause.

What survived

ICOT did not simply stop in 1992. A two-year follow-on project began in April 1993 to make the main software run on commercially available computers ( ICOT's letter), and ICOT released that software as ICOT Free Software, with copyright held by the institute and restrictions on use removed, as the archived IFS terms explain. About 100 programs were selected for release, and KLIC, a system that compiled KL1 into C, let them run on ordinary Unix parallel machines, according to the ICOT archive's conference page. In 1994 ICOT told its partners that it would close its research center once the follow-on ended in March 1995 ( ICOT's letter). Two of the parallel machines, PIM/p and PIM/m, are preserved at the National Science Museum in Tokyo ( IPSJ).

The intellectual legacy sits mainly in logic programming. Nilsson quotes Evan Tick, who worked at ICOT, crediting the project with major contributions to the implementation and theory of logic programming, constraint and concurrent languages, and deductive and object-oriented databases, while doubting the wisdom of mixing technology transfer with basic research on one schedule ( Nilsson). Tick also judged the project a success in the research it engendered in similar programs abroad, naming Alvey, ESPRIT and MCC among them. Nilsson adds the journal New Generation Computing as one lasting legacy, and records the common view that most of the project's results are now of historical interest only. What did not survive was the premise: no later mainstream computer adopted logical inference as its native operation.

How to place the Fifth Generation project in AI history

The Fifth Generation project was the largest single bet on the idea that intelligence would come from fast logical inference over explicitly represented knowledge, executed on hardware designed for that purpose. Its record is unusually clean as an experiment. The hypothesis was stated in advance by Fuchi, the targets were numerical, the money was public and steady, and the results were published. On its own terms, the machinery worked: a thousand processors ran a concurrent logic language at more than the planned speed. What the experiment showed was that the bottleneck lay elsewhere, in knowledge, applications and the economics of general-purpose hardware.

The project also set a pattern for AI policy that recurs. A national government announces a long, well-funded program to lead a technology deemed strategic; competitors read it as a threat and answer with programs of their own, as Britain did with Alvey within months. Later national AI strategies invite the comparison, though Chronicle treats any such parallel as provisional rather than proven.

Two later events in this archive frame what the Fifth Generation passed over. The neural network revival marked by learning by backpropagation in 1986 drew its power from learning rather than encoded rules, the opposite of ICOT's premise. And the parallel hardware that eventually carried AI was general-purpose, as CUDA later showed, rather than a machine built for one kind of reasoning. Neither outcome was obvious in 1982, which is the most useful lesson the project leaves: an ambitious, well-run program can meet every engineering target and still bet on the wrong constraint.

References

• Kazukiyo Kawanobe, ICOT (1984). Current Status and Future Plans of the Fifth Generation Computer Systems Project.
• Kazuhiro Fuchi, ICOT (1984). Revisiting Original Philosophy of Fifth Generation Computer Systems Project.
• Takashi Kurozumi, ICOT (1992). Overview of the Ten Years of the FGCS Project.
• Shunichi Uchida, ICOT (1992). Summary of the Parallel Inference Machine and its Basic Software.
• Electronic Computer Basic Technology Development Promotion Committee, ICOT archive (1993). Final Evaluation Report: attainment of original goals in parallel symbolic processing.
• Electronic Computer Basic Technology Development Promotion Committee, ICOT archive (1993). Final Evaluation Report: attainment of original goals in knowledge information processing.
• Hiroichi Hiroshige, ICOT (1994). Termination of the Memorandum between ANU and ICOT.
• ICOT (1993). Outline of ICOT Free Software.
• ICOT archive (1992). Outline of 5 modules of PIM.
• ICOT archive, AIST (1995). FGCS International Conference proceedings and related materials.
• Information Processing Society of Japan, IPSJ Computer Museum (1982). Start of the 5th Generation Computer Project for Knowledge-Based Information Processing.
• Computer History Museum (1981). Proceedings of International Conference on Fifth Generation Computer Systems.
• Alvey Committee, Department of Industry (1982). A Programme for Advanced Information Technology: Background and Introduction.
• Peter Marsh, The Christian Science Monitor (1983). Britain weighs advanced computer project.
• Edward Feigenbaum and Pamela McCorduck, Addison-Wesley (1983). The Fifth Generation: Artificial Intelligence and Japan's Computer Challenge to the World.
• Edward Feigenbaum and Pamela McCorduck, Creative Computing (1984). The fifth generation: Japan's computer challenge to the world.
• Ehud Shapiro, Communications of the ACM (1983). The fifth generation project: a trip report.
• Edward Feigenbaum and Howard Shrobe, Future Generation Computer Systems (1993). The Japanese national Fifth Generation project: Introduction, survey, and evaluation.
• CBR Staff Writer, Computer Business Review via Tech Monitor (1992). Japanese Fifth Generation project calls it a day.
• Nils J. Nilsson, Cambridge University Press (2009). The Quest for Artificial Intelligence, Chapter 22: The Japanese Create a Stir.
• Nils J. Nilsson, Cambridge University Press (2009). The Quest for Artificial Intelligence, Chapter 23: DARPA's Strategic Computing Program.
• Nils J. Nilsson, Cambridge University Press (2009). The Quest for Artificial Intelligence, Chapter 24: Speed Bumps.