|HAL's driving problem|
PlansIn the short term we hope HAL will play the role of a test bed for the idea of completely embedding a genetic algorithm in a cellular automata - enabling us to find the best way of implmenting such a scheme.
In the medium term we hope HAL will be able to display the ability to evolve evolvability, including the ability to evolve sex. Evolving evolvability will be critical as we are dealing with the domain of computation universal cellular automata - and these can often be rather brittle.
If higher-level representations of the target problem spaces do not develop within the model, HAL will remain confined to solving simple problems, a domain which it is neither targetted at nor suited to.
In the long term we hope HAL will become viable as a design tool for evolving electronic circuitry, which wil be able to be used as custom components in future computers employing areas of programmable logic, and be suitable for use as stand-alone components in conventional chip-designs.
AimsWe find it helpful, as a method of maintaining a clear direction, to aim consistently at a long-term goal. For HAL, this is the development of artificial intelligence in a particular, narrow domain. The chosen domain is the firld of computerised players of strategic games, in particular Go, the national game of Japan.
Go is a game with very simple rules and yet extroadinarly complex dynamics which has the virtue that it is simple to score automatically.
It is currently widely believed that Go is well beyond the range of genetic approaches. While essentially, we agree, it seems to us that this is unlikely to remain the case indefinitely and that, one day, the best Go players will consist of largely artificial evolved organisms.
We hope that the first such players will be HAL's distant descendants.
Our Go-playing program is called Gozilla.
There are some Go and computer-go programming links available.