www.gp-field-guide.org.uk
Contents Top Previous Next

Chapter 14: Conclusions

In his seminal paper entitled "Intelligent Machinery", ? identified three ways by which human-competitive machine intelligence might be achieved. In connection with one of those ways, Turing said:

There is the genetical or evolutionary search by which a combination of genes is looked for, the criterion being the survival value. (?)

Turing did not specify how to conduct the "genetical or evolutionary search" for machine intelligence. In particular, he did not mention the idea of a population-based parallel search in conjunction with sexual recombination (crossover) as described in Holland's 1975 book Adaptation in Natural and Artificial Systems (Holland1992, second edition). However, in Turing's paper "Computing Machinery and Intelligence" (?), he did point out:

We cannot expect to find a good child-machine at the first attempt. One must experiment with teaching one such machine and see how well it learns. One can then try another and see if it is better or worse. There is an obvious connection between this process and evolution:
Structure of the child machine' =Hereditary material
Changes of the child machine' =Mutations
Natural selection' =Judgement of the experimenter
In other words, Turing perceived that one possibly productive approach to machine intelligence would involve an evolutionary process in which a description of a computer program (the hereditary material) undergoes progressive modification (mutation) under the guidance of natural selection (that is, selective pressure in the form of what we now call "fitness").

Today, decades later, we can see that indeed Turing was right. GP has started fulfilling his dream by providing us with a systematic method, based on Darwinian evolution, for getting computers to automatically solve hard real-life problems. To do so, it simply requires a high-level statement of what needs to be done and enough computing power.

Turing also understood the need to evaluate objectively the behaviour exhibited by machines, to avoid human biases when assessing their intelligence. This led him to propose an imitation game, now known as the Turing test for machine intelligence, whose goals are wonderfully summarised by Samuel's position statement quoted in the introduction of this book (page  3 ). The eight criteria for human competitiveness we discussed in Section  12.3 are essentially motivated by the same goals.

At present GP is unable to produce computer programs that would pass the full Turing test for machine intelligence, and it might not be ready for this immense task for centuries. Nonetheless, thanks to the constant improvements in GP technology, in its theoretical foundations and in computing power, GP has been able to solve dozens of difficult problems with human-competitive results and to provide valuable solutions to many other problems (see Chapter  12 ). These are a small step towards fulfilling Turing and Samuel's dreams, but they are also early signs of things to come. It is reasonable to predict that in a few years time GP will be able to routinely and competently solve important problems for us, in a variety of application domains with human-competitive performance. Genetic programming will then become an essential collaborator for many human activities. This will be a remarkable step forward towards achieving true human-competitive machine intelligence.

This field guide is an attempt to chart the terrain of techniques and applications we have encountered in our journey in the world of genetic programming. Much is still unmapped and undiscovered. We hope this book will make it easier for other travellers to start many long and profitable journeys in this exciting world.

If you have found this book to be useful, please feel free to redistribute it (see page   iii:  ). Should you want to cite this book, please refer to the entry for (Poli et al., 2008) in the bibliography.


www.gp-field-guide.org.uk
Contents Top Previous Next