The hypernetwork architecture: A hierarchical molecular interaction model of biological information processing.

摘要

A novel architecture for machine learning, the hypernetwork architecture, has been specially designed and implemented in this study. This is a multi level, vertical model of a biological information processing system that includes the flow of information and feedback regulation control inspired by biological systems. The levels considered are the molecular, cellular and organismic.; The molecular level consists of many molecules, i.e., binary string representations derived from enzyme-like structures. Each molecule has an excitatory and a catalytic site, but each has an optional inhibitory site. A molecular interaction, binary string matching, represents a biomolecular self-assembly process. Dynamic formation of networks of molecular interactions represents reaction cascades in biological cells.; Molecules are placed in cells modeled by cellular automata, and an organized group of cells forms an organism. Cell to cell interactions are produced by effector-receptor molecules of the cells.; External influences on the receptor molecules of input cells dynamically trigger cascades of molecular interactions inside the cells of the organism. Then the cascade activates readout molecules on the output cells to form the output of the cell.; Hypernetwork organisms learn classification tasks by means of a variation-selection algorithm based on molecular evolution. Each iteration consists of an organism being reproduced with random molecular mutation, and the better one being chosen to perform the task. The mutation-buffering capabilities of the hypernetwork allow to search for optimal peaks in the fitness landscape.; The hypernetwork effectively learns classification tasks such as the (4–10)-input parity problem, the tic-tac-toe endgame problem, the two x two bit multiplier truth table, and a type of double spiral data set. Experimental results show that, in the hypernetwork architecture, learning improves when molecules exhibit inhibitory sites. This improvement is the result of molecular inhibition and negative feedback regulation inside the cells.