Designer chips

Citation

Rudall, B.H. (1998), "Designer chips", Kybernetes, Vol. 27 No. 4. https://doi.org/10.1108/k.1998.06727daa.003

Designer chips

Designer chips

There have always been computers that have assisted in the design of new machines and even reports of computers than can automatically design and produce new versions of themselves. Now we are told that chips have been developed that design themselves. Researchers are now said to be using the principles of evolution in processor logic design.

At the University of Sussex, Brighton, UK a researcher, Adrian Thompson, is reported to have found a way of allowing microprocessor hardware to use the principles of Darwinian evolution to create its own designs. As a result, we are told, he does not understand how the circuits work, only that they are far more efficient than any human could design.

The work is being carried out at the University of Sussex's Centre for Computational Neuroscience and Robotics. At the basis of the development was the researcher's idea to produce a breed of "self-designing" circuits while actually thinking about a project to build neural network chips. The production of efficient neural chips for networking has taxed artificial intelligence workers for nearly a decade. AI researchers have simulated the human brain with networks of artificial neurons either in software form or as "hard-wired" networks on chips.

In an interview recorded in the Computer Bulletin, Vol. 10 Part 1, 1998, pp. 18-20, and conducted by Clive Davidson for the "Leading edge" feature, Adrian Thompson explains that either way existing techniques imposed a biological design on a completely foreign medium. The human brain, the article said, has evolved over millions of years, taking advantage of the inherent physical, chemical, electrical and other properties of organic matter. What if, it stated, we could re-run evolution, only this time in silicon? Would it produce a design customised to the properties of the silicon? The problem was how to introduce the wet, vital forces of evolution, with its processes of mating, mutation and reproduction, into the dry mechanical world of silicon. To do this the researcher found his solution in genetic algorithms and field programmable gate arrays.

Genetic algorithms, it will be recalled, are programs that simulate evolutionary processes to "breed" the best solutions to a problem where there are a number of possible solutions. This is done by the developer expressing a set of arbitrary solutions as strings of binary code (the initial population). Then the algorithm checks each member of this population to test its suitability for solving the proposed problem (the fitness test). It deletes those that do badly and makes copies of those that do well. Some of the solutions are allowed to interchange parts of their code strings (mating process). In this process some mutation is introduced by allowing "random switching" of bits in the population of strings e.g. from 1 to 0 and vice-versa. The whole cycle of fitness test-mating-reproduction is repeated with the process ending when new iterations do not offer any improvements in the evolved solution.

While genetic algorithms could provide the means of simulation of evolution the problem of transferring this process to the silicon world still remained to be solved. This solution it seems appeared when a new generation of programmable chips came on the scene. This was because conventional microprocessors have their logic circuits hardwired onto the silicon. But field programmable gate arrays (FPGAs) have an array of components and the designer is able to specify the logic and the interconnections in software before downloading them onto the chips.

It therefore became possible to make a genetic algorithm create the circuit designs in software, and have the opportunity of running the "fitness tests" on the physical circuits. His subsequent experiments using this theory proved to be successful. The FPGA manufacturer Xilinx Development, Edinburgh, UK who initially provided a new FPGA that could be repeatedly reprogrammed and had no limit on what circuits could be implemented, are now the sponsors for the project at the University Centre.

Applications of the methodology are already in evidence. These include:

• •

  traffic controls ­ controlling flows of traffic such as telecommunications signals is already of interest to British Telecommunications (BT);

• •

  space industry ­ where more compact devices are needed.

Although the development is attracting much attention, issues such as the robustness of the circuit designs have to be addressed. In particular, how robust are the circuits to changes in their environment? Present research is over a temperature range of around 10°C and when this is raised or lowered the circuits malfunction. New experiments are being carried out to create circuits that are more tolerant to temperature changes and to hardware faults and the "fitness" test changed accordingly. The article says that:

There is no reason to think that evolution cannot find a strategy that would handle a wide range of temperatures. Evolution is particularly good at creating fault tolerant systems and this could be one of the major attractions of evolved designs.

Even the solution of the robustness problems may not be enough to satisfy sceptics who ask: How acceptable is a safety-critical component of a system if it has been artificially evolved and nobody knows how it works? That is the view of Dr Inman Harvey, a research fellow in evolutionary robotics at the University Centre. The answer we are told is that:

it will mean accepting that we already rely on systems designed by conventional means that we cannot prove will work under all conditions and in the rigorous testing of the evolved devices.

Cyberneticians will have already recognised the parallel with the application of Darwin's Theory of Evolution in many other contexts involving computer technology. Models of evolution, for example, have been structured on computer systems where human genetic traits have been represented as strings for a whole population. The strings are subjected to random processes, for example diseases, climate, changes, etc. and then a "fitness" test followed by a "mating" process and a "reproduction" stage. Computing machines lend themselves to such cyclic processes as seen in the numerous applications of iterative technique to numerical problems. There would appear to be no reason why such a process involving "evolved hardware" should not herald new processes that will lead to the production of new generations of machines of all descriptions. Whether we will finally be able to tell how such machines work is currently worrying researchers like Adrian Thompson of Sussex University. Whether we need to know is, of course, another question. We note, however, that evolution in the human species has not yet ended and the future may hold many surprises concerning our own evolved state.