This dissertation investigates the evolutionary design of robust and efficient connectionist swimming controllers for a simulated lamprey. Using the neuromechanical lamprey model proposed by Ekeberg [1993] and extending the work of Ijspeert [1998] on evolving lamprey swimming CPGs using genetic algorithms, I investigate the space of possible neural configurations which satisfies the following three properties: 1) Robustness against variation in body parameters, 2) Swimming efficiency, and 3) Robustness against random variation in neural connections. These properties were chosen because they are important to both the real lamprey and its robotic implementation. After a review of the relevant literatures on lamprey neurophysiology and a detailed account of Ekeberg’s and Ijspeert’s works, I describe my reimplementation of their model. Using that model, I study the effect of variation of body scale in the performance of Ekeberg’s and Ijspeert’s controllers. The controllers interact with the mechanical model in different ways to achieve high speed swimming at various scales. To investigate this phenomenon I characterise the behaviour of the body when driven by a sinusoidal-based analytic controller. The resulting performance data reveals that the mechanical model has two or three characteristic resonances which the swimming controllers exploit. The second section of the dissertation further explores the issue of robustness with respect to variation in body scale by rewarding controllers which maintain good performance across a range of scales in the fitness function used to evolve CPGs. A number of controllers are produced by the genetic algorithm and analysed, but the effectiveness of this approach is disappointing.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.

No description available.