• Mateusz Mazurkiewicz

Evolutionary engineering

In the rubbery leather and hemp bowstrings we mastered the physics of deformation of elastic bodies. We use it to sew sturdy clothing, secure luggage, or launch projectiles. Our understanding of chemical combustion allows us to send rockets soaring across the sky, and we gained control over nuclear reactions to power our cities. The scientific process, empirically driven development of scientific theories and the critical reflection on them, enables us to creatively and purposefully transform the world around us and to reach inwards, understand ourselves better. With computers we finally subjugated a law of nature which, up to now, mostly evaded our attempts at experimental study and application. It shaped the life on earth, the proof of which we see in Galapagos birds, Mesozoic fossils, and ourselves. Yet, because of its vast timescale, we could only experience a fraction of its potential.

The evolutionary process makes simulated worlds explode with diversity and wonder of artificial biospheres. Artificial life prospers in gridworlds, games, and physical simulations. Unlike a neural network trained with mathematically derived algorithms to fit the weights to minimize loss, an agency subjected to the Darwinian heuristic advances propelled by the innate, emergent law of the universe. The evolution of intelligence isn't a classification problem or a zero-sum game reduced to mathematics and solved (or solution to approximated). Rather, the evolutionary heuristic is a direction of change emergent from the underlying laws of the world. It is open-ended and will be at work no matter how difficult we try to make it, as long as an advantageous gradual change has some means to propagate.

It is an analytic proposition that life better suited for survival and reproduction will survive and reproduce better in the conditions it is adapted to. The fittest eventually prevail, whatever the world dictates as fit. Evolution seems to work against the flow of time, counter the rise of entropy. The most optimal, organized, precise structures persevere, while the matter constituting the lesser is broken down into compounds and reintegrated into the former. Two principles drive the process. If a system appears, whether designed or as a result of random events, that is more resistant to destruction or decay than others, it will be around longer for us to observe it. Gazing into the sky, there will be more stars that don't implode immediately after creation than the proportion is created. You have a short time window to observe those, and in spacetime, there is twice as much of an object if it exists twice as long.

An experiment can confirm that. A program spawns five orbs each second, each with an integer written on it. The number determines the lifespan of the orb in seconds, after which it disappears. The distribution of numbers is uniform between 1 and 10; every number is exactly as likely. Through simulated 100 000 seconds, despite 5 being the mean of the spawning distribution, the average of numbers visible is around 7. Number one can be seen about 40% of the time, while it seldom happens that there isn't any ten.

Moving onto the second principle, if a configuration emerges in the primordial mixture that can reliably self-replicate, it will spread and dominate the environment. It is no wonder there are insects. Their population rises exponentially, making the species robust to extinction. If you kill 50% of the rapidly reproducing species, it will make no difference. The population will quickly go back to the previous level, accelerated by the abundance of resources. Furthermore, the gene pool will be dominated by the specimen most resistant to the extinction event that wiped out half of the population. Unless an unusual change occurs faster than the species can adapt, such as human-induced climate change, the species will survive, and copies with better fitness will reproduce more, the fitter of progeny eventually becoming dominant.

These are the two characteristics of what exists, dictated by the emergent law of nature, a property of the categorical fact of existence. What exists is generally either difficult to destroy or swiftly proliferating. Things that are neither usually vanish into inexistence, whether lifeforms, systems, or strategies. Whole chains of evolutions: cosmic, chemical, biological, are at work. What is left is the most holistically adapted to its immediate objective reality, not one aspect or process left out of design.

This reasoning makes one appreciate a certain type of fatalism. A tiny, gentle bird species looks fragile. It has to fear not only the predator's hunt but also his careless step. Its brittle, hollow bones would be crushed under his paw. But the fact of its existence in its ecosystem is a testament to its ability to survive. The species might yet fail the test of time, and the bird might soon go extinct. But it hasn't until now, and that is a very telling fact. If an institution is on the very verge of collapse for the past twenty years, through a natural disaster and two major geopolitical shifts, it is probably not really on the verge of collapse. It might be doing awful, but the history is empirical evidence of the institution's stability, its resistance to collapse even when disturbed off balance. A pair of lovers cry at each other insults and threaten immediate breakup, but the observation that they have done that for the past four years should justify adjusting the probability of the claimed immediacy.

It was my original drive to learn computer programming years ago to simulate worlds filled with wildlife and watch advantageous traits propagate in simulated lifeforms, generation to generation. The simulations were, of course, rudimentary. I still found them wonderous. In the very first one, the agents, rectangles moving around the screen, had a gene determining the probability of reproduction. The higher that value, the redder the rectangle was. I felt like Archimedes when he ran naked through the streets of Syracuse as the simulated world was soon dominated by deep shades of red.

What I found most fascinating in my attempts at empirically exploring the topic is the evolution of emergent behaviors. What behaviors does a flock of boids evolve to navigate around obstacles? The result was a beautiful coordination. In terms of a parameter-space search problem, a boid model was optimized with a genetic algorithm. But one of the simulation outcomes was a hybrid flock consisting of two distinct groups resulting from speciation, each with different behavior. The symbiosis produced a unique flock movement.

This exploration can reach much further, from pheromone use in simple alife to emergent languages, norms, and institutions. Which was first, the human voice or the human language? The human physiology, the vocal cords and the cortex, evolved along with the development of language and culture. It was reciprocally teleological. There would be no use in evolving neural circuits recognizing patterns in speech if no one was talking. Humans built social relations appropriate to their social intelligence and emotions and evolved social intelligence and emotions appropriate to their social condition.

A point of interest in the evolution of social traits is convergence. A sense of fairness was observed in many animals, even non-mammal. Corvids reject unfair trades. How universal is it for life, and what conditions cause it to develop on the innate, biological level? Is it a development necessary to progress into a civilization, or one contingent on the particular features of life on earth? Empirical data from simulated neuroevolution can give insights into the importance and influences of ancestral environments, from the environmental ones, such as a caloric advantage of cooked meat (therefore further incentivizing being intelligent enough to start a fire), to social ones, such as pre-existing social norms or the necessity of cooperation while hunting woolly mammoths. Those insights can be incorporated into designing training for any AI to incentivize the development of social intelligence or intelligence that generalizes well in real-world scenarios.