WBL described the work on using GPU computing with finite concept lattices to find good module clusters in ten software engineering problems. This had been presented [1] the previous day at the CLA-2011 conference in Nancy.
Sebag described recent work in Machine Learning and gave pointers to "Covariate Shift" and other techniques.
WBL described the work on evolving CUDA code inside the gzip Unix file compression utility [2] and more recent experiments using Blast and related Bioinformatics tools.
Since NCBI's Blast was first created, something in excess of 140 software tools have been developed by hand which perform some kind of bioinformatics sequence matching. Blast is accepted as giving the "correct" matches between sequences, however it is too slow and resource intensive for use with next generation genome sequences.
Bowtie is probably the state of the art short DNA sequence alignment tool. It is open source and written in C++. Although not as comprehensive as Blast, it is considerably faster. WBL described plans to use it as a demonstration of using genetic programming to find new trade-offs between functional and non-functional performance.
Shoenauer agreed it was sensible to look first at functionality versus speed trade-offs. He suggested work he and Maarten Keijzer had done as part of EO and suggested "tiny CC" might be useful. However he was interested to learn in recent UCL experiments an interpreted (rather than compiled) GP system had processed 200 106 primitives per second. This lead to discussions of timing required when potentially using GP to learn optimal payoff in multi-arm bandit problems and then to how TAO are using k-armed bandits in the best computer GO program MoGo and how it might be used in optimising power supply.