Speeding up MuPlusOne
|has title::Accelerating the Online and Onboard Evolution of Robot Controllers|
|Master:||project within::Computational Intelligence and Selforganisation|
|Student name:||student name::Atta ul Qayyum Arif|
|Second reader:||has second reader::Gusz Eiben|
One of the urgent concerns in Evolutionary Robotics is to let robots achieve adaptation in operational phase i.e. online evolution. If the candidate controllers are to be evaluated using physical robots in real world, then this should be conducted in real time. However, the evolution of complex behavior will take a prohibitively long time to evolve the desired behavior. Onboard evolution makes this problem even more challenging since it requires variation, selection and evaluation to be carried out entirely on the physical robot. Therefore, it is important to devise such algorithms, which could give rapid convergence in less amount of time.
In this thesis, the author explores different schemes to accelerate online and onboard evolution of robot controllers. This thesis contains a brief literature review of different acceleration techniques in Evolutionary Algorithms. The main goal is however, to accelerate the (mu+1)-Online Evolutionary Algorithm, which gave promising results for online evolution of robot controllers. Three different modifications are suggested to the underlying algorithm to improve its convergence rate. In order to test the effectiveness of those modifications, we performed different experiments for a 'fast-forward' problem. The results were then analyzed and discussed in detail. One of the schemes is found to be very successful in increasing the convergence rate of the underlying algorithm. The devised scheme does not appear to be problem-specific and may be considered as a good candidate to speed up the online and onboard evolution of robot controllers in general.