HAL's future

Road map

HAL is currently near the beginning of its lifespan. Hopefully it will see some extensive development and fully realise its potential.

Some of the things that need doing include:

  • Create some selection criteria which modify patterns on the sides of the organisms to better illustrate the potential for evolving circuitry.

  • Produce implementations of the model on progressively more appropriate hardware. Initially this will mean using patterns in the calculation layer to specify an external phenotype which can be implemented very simply and rapidly in programmable logic.

  • Rework the model so all offspring have the same orientation (rather than being pointing in four possible directions). This would removes the restriction on the calculation layers that they be rotationally symmetric and simplify the growth of cells.

  • HAL's long-term viability as a method of solving optimisation problems depends crucially on our ability to develop sexual organisms, or those that exhibit significant crossover frequency.
    Lack of sex appears to be a fundamental stumbling block which currently prevents artificial life models from competing properly with more conventional genetic algorithms which employ externally imposed sex.
    We plan to create an environment where sex evolves naturally from asexual organisms. This will involve creating a number of other models, mainly of parasitism and gene repair, to investigate what dynamics are most likely to lead to the desired outcome, and then finding methods of creating a suitable environment for the evolution of sex within our main model.
With regards to the optimisation problems we plan to use, in order:
  • The ability to solve simple arithmetic and logical problems.

  • The ability to use memory. We plan to present organisms with a series of numbers and then return them at a later date, on request.

  • The ability to form multi-cellular organisms. We plan to drive organisms into this by selecting for small organism size in addition to presenting particular organisms with harder memory problems than any individual can handle in the hope that they will delegate parts of this to their sisters.

  • The ability to play strategic games. See the details of HAL's driving problem for more details about this.

The point behind using a set of problems connected with memorising sequences of numbers is that we don't believe in rewarding organisms in our model who perform in a pleasing manner directly with reproductive success, rather we prefer to pay them for their services, and then charge them for reproductive facilities at a later date.

There are essentially two main methods of implementing such a scheme, one involves offering the organisms a precious material, or scarce vitamin, essential to reproduction in payment for their services. The other is to pay organisms for their services in e-money and then extract payments from them when they attempt to reproduce. We have chosen the second alternative over the first on the grounds that it simplifies our model.

Paying organisms with e-money allows them more choose concerning when they reproduce, provides a medium of exchange for trade within the model and encourages parasites within the model. Parasites are important for the evolution of sex. The positive effects they have in this area and in terms of driving the model away from local minima by continuously deforming the fitness landscape more than offset the problems that they cause in terms of causing organisms to "waste" resources fighting them and when finding a 'model' organism who is healthy at the end of a run.

Index | HAL | HALcell | Parallel | Hardware | Future | Problem | Limits | Automata | History | Links
tim@tt1.org | http://www.alife.co.uk/