Today is Charles Darwin’s 200th birthday, and I would like to celebrate it by drawing some connections between his theory of evolution and recent advances in machine learning.
It took the genius of Charles Darwin to break with the common belief that species were constant and unchanging. His voyage on the Beagle was essential in forming these ideas as he encountered fossils of extinct forms and the effects of dramatic earthquakes in the Andes, and was led to realize that the Earth changes and that species change. This idea of long-term change is hard to come by. It even evaded Albert Einstein when he derived the expansion of the universe from his theory of general relativity and decided to introduce a correction term in the mathematics to ensure that the universe was constant as he knew it to be. Einstein called this his biggest blunder. Yet it was exactly this belief of constancy that Charles Darwin was able to shed.
What few people realize is that Western capitalism has embraced Darwinism and use the arguments of survival of the fittest to defend deregulation in the market. This competition does work well… as long as you actually have competition.
In the area of data analysis and machine learning, which are my areas of expertise, we rely heavily on computer algorithms that search vast high dimensional spaces for solutions to problems. The best algorithms employ techniques that are central to evolution. These methods are called Markov chain Monte Carlo (MCMC) techniques, and in some specific cases these algorithms have direct relation to genetic evolution and are called genetic algorithms. What makes these algorithms work is precisely what makes evolution work.
These algorithms rely on a fitness function that enables us to measure the fitness of a hypothesized solution to a problem (we call this a sample). We often start with many samples scattered throughout the space and let them explore via mutations. In some algorithms, we take good samples and duplicate them and let them explore further, or we may even combine characteristics of a pair of samples to create a new one (as in genetic algorithms). After hundreds of thousands of iterations, the algorithms are able to find the solutions to the problem. These solutions would be impossible to find via brute force search or guessing.
The analogy that holds here is that of organism as sample. Anyone who has actually done these simulations can understand that you can obtain solutions worthy of creation simply by iterating hundreds of thousands of times under the force of a selection pressure.
Happy Birthday Charles Darwin, and thank you for the insights that have advanced machine learning in the last two decades.
Kevin Knuth
Albany NY
Posted under Biology, Evolution, Intelligent Systems, Robotics
