Genetic algorithm
A genetic algorithm (GA) captures key elements of the process of evolution to perform search through high-dimensional spaces. In the space of adaptive learning algorithms, gradient-based error-driven learning is the best, trial-and-error reinforcement learning is the second best, and genetic algorithms are the third best, in terms of speed of convergence.
This hierarchy is also defined in terms of the generality of the mechanism: error-driven learning has the strongest requirements (needing target output patterns), while reinforcement learning only requires a scalar reward signal, and genetic algorithms likewise only require a definition of fitness, without the need for the ability to compute a gradient. Thus, it is the most general, robust form of learning. But like evolution, it can take a while.
The essential ingredients of a GA include:
- • A genotype, which is a compact modifiable plan for how to build a phenotype (organism, entity, etc).
- • A fitness function that evaluates the performance of individual phenotypes.
The process of evolution thus proceeds by randomly generating genotypes, creating a population of phenotypes from them, and then evaluating the fitness of the individuals in the current population. The top-performing subset is then “mated” to form a new set of genotypes by mixing elements among them, and the process iterates over multiple generations. There are many variations in each of these different components, as summarized in the wikipedia page.
From a pragmatic perspective in relation to the Axon and Rubicon models, GAs are typically too computationally expensive to be of significant value. Simulating a single brain-level model of any complexity is already a very computationally expensive process, so iterating over large populations of such models for many generations is not an efficient way to proceed.
Instead, we take the approach of reverse engineering the brain through neuroscience informed by research on cognition and computation to conduct a more directed, hypothesis-driven search for the key “discoveries” that millions of years of evolution on our planet has produced.
Nevertheless, in specific narrower contexts, GA algorithms are the only way to proceed.