Created by W.Langdon from gp-bibliography.bib Revision:1.4759
This proposal is a result of research over the past two years, and whose purpose was to develop a design methodology for low cost ultrasonic blood flow and tissue quantification using signal processing. My original desire was to improve feature extraction techniques for use in statistical pattern recognition, but was almost immediately redirected along the lines of efficient genetic search of network solution spaces. Over ten years of experience with Doppler flow measurement suggests that dynamic processing of the clinical signals involved can be done with interconnected functional elements such as delays, filters, and thresholds. Some details of the processing issues and reasons for using genetic search will follow. The point of this dissertation is to study and develop a specific method for synthesising processing networks that aid in the use, interpretation, and diagnostic power of low-cost medical technology.
Results of synthesising a signal processing network that correctly recognises fiducial points in a simulated two-heart cycle, spectrally represented, wave form suggests the ability to handle similar applications with real clinical Doppler data. The solution described in the previous section made use of a delay element that matches the heart-cycle period and is otherwise sensible. Search difficulty was increased by including in the function set a number of function/operators not actually needed to solve the problem. This was done purposely to eliminate the necessity of defining a problem dependent function set as may be necessary for medical data.
A multiple trial, multi-modal, partially deceptive test problem provide further evidence that the Max(f1,f2) diploid/dominance implementation can provide better than or equal processing efficiency, compared to haploid. This conclusion is supported by a similar, though less thorough, comparison using the R-wave network synthesis problem. The Max(f1,f2) approach has been observed to do about the same as haploid with either very simple (e.g., unimodal) or very difficult or poorly formulated problems. Diploid/dominance as implemented here can be used in conjunction with other improvements (e.g., more refined crossover, inversion, species formation, etc.) to the standard GA. The experiments with alternating fitness environments show that multiploid populations are capable of storing and rapidly recalling as many global optima as there are homologues in each individual chromosome and shows that diploid/dominance retains recessive alleles and schema.
The diploid approach could immediately make use of a two-processor system, since the algorithm used involves two function evaluations per generations.",
Genetic Programming entries for Francis Greene