Thoughts on artificial life, death and intelligence

I’d planned to post this on the 13th, just because that date would have been appropriate, but after implementing death, I was tired and a little bit sad. Combine that emotional response with this obviously guilt inspired dream, and the evidence suggests that I’m a gentler soul than I thought. It does niggle at me—the idea of creating life like bits of software and deliberately making certain that they will die. Still, it’s only after implementing death that I really consider them alive. That’s not to say that I consider death to be a necessary part of life, but in this approach, without death, there probably isn’t enough pressure to cause evolution, and without some evolution, my little subleq based creatures won’t be able to learn and grow.

There are examples from biology of organisms that don’t seem to have an inborn limit to their lifespan. The only way they die is if something kills them. I can think of several ways that evolution could happen without implementing an external death function.

The figures can and do interact with one another, and there are several ways in which one figure could “kill” another one. One figure could cut out the entire contents of another figure’s memory. One figure could modify another so that it only does one or two meaningless commands. A figure could write “0, 0, 0,” at the beginning of another one, causing it to just loop that single command. The figures can already do any of those, and any of those would keep the figure being modified from changing itself, changing other figures, or copying itself to a new slot. The figure would still be there, but as it could take no effective actions, it would effectively be dead. That’s one of several experiments I’ll be doing—to see if the figures can evolve learn and grow, without any death coming in from outside.

If I had infinite time and resources, none of them would have to die at all ever. Instead of killing figures, figures that are behaving in especially interesting or useful ways could be copied to a different population, and allowed to continue from that point. It would be a form of selection that didn’t rely on deleting anything. In some ways that would be ideal. It would mean that each level of the population would be continuously feeding into the next level. That would provide a much wider search across the hills and valleys of the fitness landscape. However, I don’t have infinite resources, and it would require that I create a fitness function to act as the selection mechanism, and that somewhat defeats the purpose of using digital organisms instead of an algorithm like genetic programming.

In a way, I’ve been working up to doing this project since I started this website and podcast. I didn’t know that it would be this project particularly, but much of the show has been a study of the mind, including the universe that it’s imbedded in, and I’ve had several possible experiments with artificial intelligence bubbling around in the back of my overheated brain.

I’m dissatisfied with the results of other methods of bottom up AI. They lack the capacity for independent innovation. They can generate all sorts of useful answers to all sorts of interesting problems, but they don’t spontaneously come up with novel behaviors and strategies outside of the problem domain they’ve been designed for. unlike actual minds and biological systems, they only work on the task you point them at. They’ve also, thus far, shown little to no capacity to generate a general intelligence—a mind that can come up with reasonable answers to multiple problems of arbitrary complexity.

It’s possible that the limitations are a matter of how large, or small, the implementations are. Even the largest neural networks, for example, are only just scratching the edges of insect sized brains, and the algorithms are based on an extremely simplified model of a biological central nervous system. Perhaps, as we are able to implement larger and more complex models, they could pass some threshold, and start to exhibit some of the behavior that seems to be missing; but there’s another problem that this project is meant to investigate.

With both evolutionary approaches, and neural networks, how well the system is performing, the fitness function, is designed by us. That means that our preconceptions get built into the algorithm. There may also be an over emphasis on efficiency. Different approaches are regularly measured against one another, and all the attention—and funding! —tends to go to whichever approach seems to take the least amount of time and computing. Compare that to nature, where several mutations that are a disadvantage when they happen in isolation can come together and end up being a strong advantage when they occur in concert. Lastly, we’ve the lingering effects of Darwin’s original theory, with its winner-takes-all assumptions. It’s obviously from even a cursory glance at nature that you don’t have to be the very best for you and your progeny to survive and thrive; but those ideas and attitudes still pervade our thinking, in realms from AI, to business, to finance, to government, and even personal relationships.

That’s why I chose to start with artificial life, and have been developing a Tiera like system. Digital organisms regularly develop their own fitness functions, and the interactions between members of a given population become at least as important as any externalized evolutionary pressures. I’ll still need to provide fitness functions when I want them to work on a problem, but it is my hope that the figures’ interactions with one another will help keep them from over fitting. I’ll be experimenting with using the same population to work on wildly different problems, just to see what happens, and I’m not going to be overly concerned with how efficiently they perform when compared to other approaches. I want a general intelligence, or at least some glimmer of a hint of one.

Meanwhile, as a way to soothe my troubled conscious, I could think of the overall system as the thing that is alive, and the birth and death of each individual figure being analogous to the life cycle of individual cells within a multicellular creature. Squint at it right, and it’s even true.