Returning again to the FoldIt game. Here’s the computational problem: finding the solution is really really hard. Okay, more precisely, there are multiple gazillion different potential solutions to be tested and even with uber-fast computers, to examine the entire ‘search space’—that is, test every last possible solution to find the best— would take more time than any of us have. Much more. So. This means in reality only a small subset of possible solutions can be tested. Think of the possible solutions as represented in a multidimensional space. What areas of this space should be explored?
Enter heuristics, or ‘rules of thumb’ used to guide selective searching within an impossibly large space of solutions. So computer algorithms are designed to search only solutions that could, possibly, make some sense. They are also designed with particular rules for selecting the next possible solution depending upon the results of the last, resulting in a ‘trajectory’ or sampling method through the search space. Here’s where it gets interesting. Humans show a much more varied set of search strategies and skip around a much larger range of the solution space, even venturing into ‘ridiculous’ territory where the solution clearly couldn’t be correct, areas a smart machine would avoid altogether. And yet. . . in this merry dance through the solution space, sashaying through ridiculousness, the humans arrived at a better solution than the machines. What’s brilliant, is that the different heuristics developed by humans can be deployed to machine algorithms, possibly giving the machines a leg up and tightening the competition a bit for future rounds.
What’s interesting here is that heuristics function as a sort of filter, a lense through which to view the world, a way to reduce the complexity of the world and see only those aspects of it that are meaningful and actionable. Such filtering of the world and information is critical for efficient processing, whether machine or man. But it’s a double edge sword. Such filters and heuristics can be thought of as habit. Given any random space of stimuli or information, we apply heuristics, based on previous experience, and see the world (or process information) in a habitual way. Though such habitual ways of seeing and thinking increase efficiency, they can also function as blinders. These habits, or ways of seeing the world, are programmed into machines: heuristics. Aspects of the world outside of one’s habitual view often become essentially invisible. That’s what happened to the computer algorithms; certain regions of the solution space were effectively non-existent. People also acquire heuristics, or habitual ways of seeing and thinking, over long years of experience. In FoldIt, though, most of the players had no heuristic way of viewing the search space as it was entirely new to them. Critically, they were not taught all the rules and methods instilled in software algorithms. They came to the task, and the search space, free of habit, free to roam the entire space, even those ridiculous regions essentially invisible to machines.
It’s tempting to muse on the superiority of the human mind, but that may be missing the essential challenge faced by both man and machine: time is of the essence and more often than not you have to reduce the computational load by trimming the world and/or solution space down to manageable sizes: without having the opportunity to evaluate what you are cutting out. Think about that for a moment. Carefully.
Deep Blue beat Kasparov in 1997 (unless you think IBM cheated), but the computer was processing 200 million possible positions per second. I am absolutely certain Kasparov’s brain was not doing that. Kasparov filtered the game space—and the vast number of possible move sequences— through his experience, knowing in advance, as it were, those regions of ‘move space’ that need not even be considered. Deep Blue lacked these exquisite heuristics and applied brute computing force to the problem. Even if Deep Blue won, it lost big points for efficiency. It’s doubtful it could have survived any sort of evolutionary challenge. IBM dismantled it as a mercy killing.
So in Foldit, once again we seem to have human idiosyncracy and intuition seeing the world in unusual but, as it happens, remarkably efficient ways. And, ultimately, teaching machines to be more efficient. I wonder, though, whether human genius and efficiency is self-limiting. As FoldIt players get better and better at the game and become more and more expert, will they themselves develop more entrenched heuristics, ways of seeing and approaching the problem? Will the solutions to some proteins elude them because they fall in a part of the solution space that has become, over time, invisible to them? And if so, will newer players, free of the blinders of habit and heuristic rush in and develop alternative heuristics and find those recalcitrant proteins hiding in the solution space safe from both computer algorithms and seasoned FoldIt players? I really don’t know. It will be very interesting to see.
A last thought: the one thing that might save human players from becoming locked in habit and blinded to parts of the solution space might be social interaction. Something computers don’t have. Yet. Social interaction provides a way out, a way to not become mired and blinded by one’s own heuristic algorithms. I think there lies the fundamental difference between man and machine and, in my estimation, the computational power of the FoldIt players.