Genetic synthesis of signal processing networks utilizing diploid/dominance

A pattern recognition methodology is presented for synthesizing signal processing networks, which are used to solve a low-cost medical signal processing problem. The approach makes use of genetic algorithms and a new approach to diploid/dominance, which is tested using both artificial and clinically obtained data. Networks are expanded from the genotype using a grammar that is based on node arities, which is in some respects similar to genetic programming (GP). Maintaining a population of genotypes provides for separation of genotype and phenotype that is absent with GP, and also permits future implementation of powerful mechanisms found in nature, such as self-assembling gene products. The effect of diploidy on search efficiency is analyzed and evaluated with three test problems. These are: (1) A previously used 0-1 knapsack problem, used to test diploidy with non-stationary fitness criteria, (2) A fast to execute and easy to replicate multimodal test problem and (3) Multiple trials with a network synthesis problem using clinical Doppler signals. The diploid implementation is completely problem independent (e.g., can be used with GP) and appears to provide some efficiency improvement with non-stationary fitness criteria. Two complementary reasons for the observed diploid efficiency increase are proposed: (1) Improvements due to the retention of relatively low-fitness; recessive building blocks and (2) Improvements due to the increased proportion and fast evaluation of non-viable, recessive genotypes.; The network synthesis approach makes use of a large number of network function primitives, when compared to previous uses of GP. These functions include both simple, general purpose functions such as arithmetic and Boolean operators, and more complex, problem specific functions. The problem specific functions are designed to take advantage of known or intuitively desirable signal measurements. These include an adaptive filter, noise resistant sideband estimation and band limited spectral energy measurement. Parameters for the nodes are taken from the genotype and can thereby evolve. The demonstrated use of a relatively large function set promotes design flexibility and increases the adaptability of the approach to other problems. The ability of the designer to suggest a trial solution through genotype "back" coding is also demonstrated.