Genetic algorithm takes over experimental control
Experiments with ultra cold atom and Bose-Einstein condensates are highly complex, containing hundreds of input parameters (usually handled by computer-based control systems), that a physicist needs to optimize to make the experiment work. Interestingly, the outcome of a measurement is rather simple (mostly a photograph of the atomic cloud) and the few crucial parameters can be derived automatically and in real time.
We have now successfully implemented a genetic algorithm to automatically operate and optimize our experimental setup without human intervention (link to article here?). Genetic algorithms use concepts from evolution. They start with random parameter sets and only the ones that perform successfully in the experiment "survive". The "fittest" parameters are combined to "breed" to new parameter generations, "mutation" introduces a random element to keep population diversity and so forth until the system finds the optimal operation conditions.
Genetic algorithms have been applied in various fields, notably in evolution research, informatics, but also in game theory and in chess or GO programs. However these applications fulfilled always highly specialized tasks.
Our aim is to show that today, where many research experiments are computer controlled, and the analysis can be performed on-line by computers, genetic algorithms can close the loop between input parameters and the experimental outcome to help the scientist in the time consuming tasks of optimizing the experiment, to find the best 'working point', where the experiment performs best and most stable.
Besides being intellectually intriguing, we are convinced that this new approach will be widely used in the future to realize "experimental optimal control" in analogy to optimal control theory in numerical simulations of physics experiments. Our approach is not limited to physics experiments, but can be applied to any complex experimental procedures which requires so many diverse inputs that standard pathways to optimize are too time consuming and impractical. The generic algorithm approach will bring substantial help in making the experiment work and create a stable experimental environment.
Literature:
W. Rohringer, R. Bücker, S. Manz, T. Betz, Ch. Koller, M. Göbel, A. Perrin, J. Schmiedmayer, and T. Schumm
Stochastic optimization of a cold atom experiment using a genetic algorithm
Appl. Phys. Lett. 93, 264101 (29 Dec 2008), doi: 10.1063/1.3058756
Mark Buchanan
Fit for purpose
Comment in Nature Phyiscs 4, 901 (2008), doi: 10.1038/nphys1138
