Browsing through material on genetic computer methods I have on several occasions encountered a very interesting phenomenon – genetic methods that exploit flaws in their environment to help achieving their goals.

Using genetic methods, people try to “breed” software that best meets the task that is to be solved (the goal), just like nature’s evolution processes breed animals that are best fit to survive in their natural environment. In this process nothing is “forbidden”, the individuals can try whatever means available to them to seek the optimal solution, and in the case of a computer environment that can mean exploiting unknown or at least not-intended-to-use flaws in the environment’s design.

This spring I wrote an essay called “A mind emerges” in a Philosophy class. In the essay I mentioned this phenomenon and used two examples:

1. Hardware chip flaw: In an experiment to train a neural network to distinguish between different frequencies of sound, Professor Adrian Thompson of the University of Sussex ran into an interesting problem. When he had trained the network and was satisfied with its accuracy rate (which was almost 100%) he took the network that had evolved and started to deduct the parts of the network that did not contribute to the task of distinguishing the frequencies. Using various methods, he removed parts of the network and simplified it.

   But at one point something unexpected happened. Removing a link that did not seem to contribute anything to the real functional circuitry made the program fail. The link wasn’t even a part of the circuit between the network’s input and output! Obviously, this did not make any sense at first, but after doing some further experiments he found the reason. The computer chip he was running the neural network on was flawed. When the network that he had successfully trained on this particular chip was copied on to another chip that was supposed to be identical to the previous one, the network didn’t work at all. When training the system, the network had found a way to use the flaw in the chip to its advantage.

2. Software physics model flaw: Another similar story comes from a work by GenArts genius Karl Sims on simulated creatures that were evolved in a simulated environment. Sims used genetic methods to develop means of locomotion for his computer creatures (not unlike the snake). But a few of Sims’ creatures developed a very strange and what whould under normal circumstances have been impossible means to move. This turned out to be a result of a flaw in the implementation of the physics in Sims’ simulated environment. The evolving creatures had spotted the flaw and exploited it to their advantage. (Discussion on this exploit can be found in ‘Frontiers of Complexity’ by Peter Coveney and Richard Highfield, see: Spreading the Cause).

Both of the above samples are very interesting and are good examples of this exploiting behavior. Now what I’m wondering about is what this tells us about methods and solutions that have evolved in the “real” nature? The fact that animal organs sometimes evolve to serve a different purpose than before seems to be of similar nature. Exploitation of this sort certainly leads to more complex organisms rather than simpler, as a minor “defect” in an individual could be exploited to serve as an organ with a specific purpose. And what might this mean for as complex systems as the human brain? Without doubt helpful defects on the evolutionary road have been used or as I put it in my aforementioned essay:

“[Exploiting of this sort] shows that a neural network cannot be abstracted from the environment in which it functions. We cannot necessarily simulate the functionality of the brain just by simulating the neuron functionality. Answers might be found in the neural network’s environment, e.g. the glia cells (that are mainly believed to be a support system for the neural structure) or even in how the skull affects the brainwaves or how the blood flowing through the brain carries different chemicals. This might mean that even though the brain’s neural network carries out the main functionality of the brain, nature might have found ways to exploit “defects” in its own design for its benefit in a similar way that Thompson’s neural net exploited the flawed chip or Sims’ creatures exploited the flaw in the physical simulation.”

Links:
Karl Sims’ Evolved Virtual Creatures
Video of Sims’ creatures (don’t miss this one, it is fantastic)