Design Challenge 6

Challenge Details

In the Evolve a robot that can draw Design Challenge, reproduction takes place via crossover - where two individuals genomes are spliced together (corresponding to sexual reproduction) - and mutation - where one or more members of the genome are randomly altered by a small amount.  In terms of the solution space (or fitness landscape) mutation enables hill climbing that optimises behaviour with the obvious risk of becoming trapped in a local maxima. To overcome this limitation crossover allows valley bridging and it also attempts to ensure that the solution space is sampled as effectively as possible (within the limitations of the initial population).

The system can be seen as an n-dimensional chaotic system where the fitness functions define convergent attractors that are capable of expressing the desired outcome or behaviour.

The design challenge is to identify fitness functions that are requisite to the task of evolving autonomous/independent drawing behaviour.  It the fitness functions are too “high” it is probable that the drawing behaviour will emulate the drawing styles of the people who have defined these functions.  It should also be noted that genetic algorithms are extremely exploitative and poorly defined fitness criteria are likely to result in unexpected and inappropriate outcomes.

It is proposed that sets of “low level” fitness functions exist (that may be associated for example with the fundamental mechanical actions of pen movement -or- mark interactions) that will enable the desired behaviour to evolve.  The task of this challenge is to identify such functions or their alternatives or maybe possible routes to their future discovery.

As readers may be aware this maps onto a “real” design challenge and was recently funded by the AHRC as “Computational Intelligence, Creativity and Cognition: a multidisciplinary investigation” - see the URL:  http://www.cogs.susx.ac.uk/ccnr/research/creativity.html