This first entry should explain the hows and whys behind this obscenely weird and long blog name. In brief hypothesislessness describes the state in which we as scientist sometimes find ourselves when we are facing a new problem nobody else worked on before.
This status became very apparent to me in a very special meeting of the Evolution of Intelligence group here at BEACON MSU.
We address two questions: How did intelligence evolve, and how can we use evolution to create an intelligence. And we are not the only ones doing it, and we are not the only ones who think they work on these questions. I have to point out that we have a very heavy emphasis on doing actual evolution. And instead of trying to reconstruct evolutionary events and how evolution might have happened and how it shaped the beings around us in the process, we evolve things or beings and ask what parameters influence evolution and how that might be used in engineering. When other scientist do their work, they normally work in a well established field, that contains tons of concepts and testable hypothesis. But I think we are in a very different situation. In most cases we study how basic behavior like motion, migration, swarming, navigation, or cooperation evolved. And quite frankly, I know nobody who witnessed any of this. Of have you been around when the first amoeba stretched here first lamelopodium? Of when the first cells started to cling together? Or maybe you know which insect first laid the egg of a worker minion? What most people do is look at a system that is, and try to answer the question "how it came about". What we do is using a system that evolves into doing something and we observe how it happened, and see which parameters influence the outcome. But there are no hypothesis about these kind of processes. There are vague ideas of how something happened in species X and that maybe factor Y was required. I am in a state of hypothesislessness, I don't know what hypothesis to test in the evolution of intelligence. I find myself in a situation where I have the tools to observe something nobody else ever looked at, and I might be able to make profound statements about the situation we are currently in and the inner workings of everything living all around us, and at the same time I get the impression that science missed this field entirely. To give an example: Bee swarming. Of cause bees evolved to swarm coming from a non swarming ancestor. And yes, scientists have good ideas how bees became what they are. But they are unable to make any kind of experiment in a biological system that even remotely incorporates actual evolution. For that you would need to make an evolutionary experiment with non swarming bees (or bee ancestors) make them swarm, and test which conditions where conducive. Good luck! It would take the time the entire humankind took from the upright first step to now to pull of this experiment. It can't be done. Or can it? We can evolve virtual critters to swarm. I have done that, it is surprisingly simple, and I can test parameters until I am blue in the face. Did I learn something about bees? Most likely yes, if I take my results and make generalizations about actual bees, I might come up with a couple of hypothesis, which will be very hard to test in the actual system. Still I can use experimental computational simulations to generate testable hypothesis. But do I find nice testable hypothesis in Biology that I can test? No, because Biologists can't make such an experiment and therefor they don't think about hypothesis testable in this manner. I want to turn this around. I want to use our computational evolutionary systems, which because they implement evolution rather than simulating them, and thus are an instance of evolution and therefor allow to test hypothesis, to actually test biological hypothesis. I don't want to generate hypothesis that Biologist might or might not be able to test, I want to test their hypothesis, and I have a hard time finding them - I am hypothesisless.
Of cause I am exaggerating. I trying to provoke a discussion. I want you to tell me that I am wrong, and that there are tons of biological hypothesis in the field of evolutionary intelligence and behavioral science I can test in my computational systems. And I am also fully aware that testing hypothesis in the best case creates new testable hypothesis. But I think we are not using our tools efficiently. Evolving some form of behavior and coming up with ideas how it might have happened in nature just to find out that we can impossibly test them in the wild frustrates me. Let's turn the board around, and let me test your hypothesis for a while.
Cheers Arend
Recently Dr. Charlotte Hemelrijk was visiting and gave a couple of great talks on her studies of flocking and schooling. She builds computational models of observed behavior to try to deduce the rules used to create the observed emergent behavior. So she is on the crossroads of biology and computer modeling. In developing her models she make a proposal (hypothesis) and implements and verifies it against biological observations of gross group behavior. Then she goes back and adjusts her model, trying to keep what she thinks are all relevant factors involved. When shown the Hyena model we evolved at the Univ of Idaho and Arend's bee swarm evolution, the first thing she wanted to know was if we specified the rules and what they were. I think she meant like in her experiments. Both Arend and I explained the open endedness of the experimental runs. But I think she was dissatisfied with the approach probably because, dare I say it, it was hypothesisless.
ReplyDeleteBut then I think not. I think this might be the difference between heuristics and meta-heuristics or X and meta-X. We can hypothesize what evolution might need to find a solution that matches biological observation. It is one or two steps removed from the cause and effect of classical biology. It is as if we asked what process would be necessary to have the behavior we observe evolve. So we replace her thinking up a model, implementing and testing it with we think up a search space of possible models and a process to search it that looks like evolution and then we try to have the machine evolve it. When it doesn't we go back and modify our model as she does and repeat. I am not sure it is all so different it is just whether the "result" is at the meta level or not.
I think Dr. Hemelrijk was also disturbed by this: when you get the evolution right and you are evolving solutions, we can look at the solutions (stripping away a meta, as it were) and now the machine has proposed a solution to say flocking. The specific solution is not general like a hypothesis or if it is it is cryptic and so does not enlighten the researcher. It comes down to: she wants to know where is the hypothesis related to observed biological behavior the biologist was interested in. That is her focus not the evolution of the rules of the behavior but the rules themselves. The questions are different. They are annoyingly intertwined and yet annoyingly separate in that one may inform us little about the other.
Dr. Hemelrijk shocked me in a different way, so to speak. After she looked at the swam she simply said: Of cause you see swarming, you selected for it! I was so perplex by her confidence in evolution that I couldn't think of a good response. All our lives would be so easy if her statement would be true. We are spending most of our time designing fitness landscapes, and getting them to show the behavior we were looking for is the hardest part. I was tempted to say: "You fiddle around with parameters until your birds swarm, which proves nothing! While we at least come up with evolved behavior." That would have been very hostile, and of cause I would never be like that.
ReplyDeleteGetting back to hypotheses: Her birds swarm and move around even without a predator, while our swarms hang out where they started, not roaming an area. We think, that predation plays a crucial role, and we sure should put a predator in this system to show: Swarms require predation to evolve roaming behavior, while unpredated swarms keep their center fixed. As you said, her approach will result in a set of rules which determine how a set of black pixels look like a bird swarm, while our approach shows how a predated set of white pixels evolves to look like a predated bird swarm, while an unpredated does not. I said above: "That proves nothing!"
But I think that our approach proves something because first: We actually make an experiment, which allows us to establish causality. And secondly, we make a statement about the process of swarming and about the reasons of swarming, not about the question "how to built something that swarms".
Several questions remain: Did any birdswarmologist every put forth the hypothesis that roaming swarms require predators to evolve? And if so, was this in the discussion mentioned as some possibility, or the actual subject, comparing predated and unpredated birds and their swarming behavior making inferences about past events? And if indeed someone published this, I bet the paper makes lots of points about how evolution happened without actually studying the process, and without ever evolving anything. For sure, we would never get one of our publications in nature, because we didn't do anything biological, purely computational. And the prejudice remains that we can do everything in computation. At the same time, we would have been the ones showing a causal relation, not only a correlation...
Cheers Arend