One of the most powerful critiques from Full’s lab is that our Tadros are too simple. Therefore, the argument goes, the Tadros are not accurate models of biological phenomena. Tadro3 fails as a model at the level of an individual agent. As with all useful critiques, it’s true, at least in part. So as prey items, eager to learn, we first try to admit that we have a problem. Then we try to figure out how—or if—to fix it. One defensive fix is to attempt to do a better job of explaining the different ways that you can judge biorobots as scientific models. That’s why I summarized Barbara Webb’s approach to biorobotic modeling in the first place.[108]
Another fix is to go back to the beginning and revisit the KISS principle. We used KISS to justify creating a simple model first. Now we were being criticized for the simplicity of Tadro3. So what’s next? What new elements do we add to the model, and why? Will those new features, which will certainly make our robots more complex, also be the right ones to make our models more accurate mechanistically and in terms of representing biological systems? To answer the what-next question is why I took the time in Chapter 5 to talk about embodied brains, neural circuits, and the subsumption approach to modeling the recticulospinal sensory-motor system of fish. We add biologically based complexity to the nervous system of Tadro3 to create Tadro4.
But what about testing a hypothesis? Is the Tadro3 system relevant? Yes. We set out to use our Evolvabots to test the hypothesis that selection for enhanced feeding behavior drove the evolution of vertebrae in early vertebrates. We substituted structural stiffness of the notochord for number of vertebrae in a vertebral column. Because the population of Tadro3 evolved reduced structural stiffness but better swimming behavior under selection, we disproved the hypothesis that selection for enhanced feeding behavior drove the evolution of vertebrae. However, we came to recognize that the hypothesis itself is probably overly simplistic, given that feeding behavior and structural stiffness of the notochord can become decoupled with a fitness function that rewards multiple sub-behaviors for feeding alone. This led us to add complexity to the selection environment of Tadro3, tossing in a predator to create the wild and crazy new world of Tadro4.
Another set of critiques of Tadro3 came from the predators within. Every time we got our multi-institutional research group together, Rob Root, professor of mathematics at Lafayette College and one of the team leaders on the evolutionary simulation project, was always itching to sink his teeth into the robotic Tadro3. One of the many great things about collaborating with Rob is that his bites are gentle and always meant to be constructive. It also helps that he is persistent when we want to deny that we’ve got any problems. As we slowly unrolled the methods and results from the Tadro3 experiments over many months, here’s the blood in the water that got Rob’s jaws clacking: (1) our fitness function was a composite of feeding-related behaviors rather than being, for example, the actual amount of food collected; (2) our fitness function was a sum of unscaled measures; (3) the composite feeding behavior did not include acceleration performance; (4) the population size was too small (remember in Chapter 2 that Rob had brought up the issue of the large effect that randomness can have with a small number of cases); (5) the number of generations was too few to show large evolutionary trends, and (6) vertebrae were missing in action.
I ran these recollections by Rob recently. He agreed that these had been his primary critiques. He had always found it fascinating, he added, that vertebrae evolved independently multiple times in vertebrates. That convergent evolution, he noted, appeared to be contingent upon the prior evolution of enhanced sensory systems, notably the paired nose, eyes, and ears that characterize vertebrates. This observation is part of the justification for the sophisticated nervous system and sensor systems that we talked about in the last chapter: Tadro4, to succeed as a potential prey in the prey-versus-predator world, needs to have the nervous and sensory-motor systems to give itself a fighting chance of survival.[109]
At the time of Rob’s critiques, all we could say, given that we couldn’t change our evolving Tadro3 system mid-stream, was, “Great points, Rob! Well taken. Next time.” Fortunately, at the same time that the robotic Tadro3 system was busy evolving, Rob and another one of our collaborators, Chun Wai Liew, associate professor of computer science at Lafayette College, were taking on the Tadro3 system and the biological hypothesis it tests using a different technique: digital simulation.
I’d done my own predatory critique back in Chapter 1, bringing up digital simulation only to dismiss it. The dismissal part was not entirely fair, I know, and it raised some hackles on my friends who gave me feedback on the early draft of that chapter. My point, which I still argue is valid, was that embodied robots have the advantage over digital ones because the embodied ’bots can’t violate the laws of physics. This is true. However, if you can build an accurate “physics engine”[110] and use it in your digital simulation, then you can avoid creating completely unrealistic models. You can even learn something too, as we’ll see.
To model a world that gives your actions predictability, you need the rules that a physics engine provides. Gravity, momentum transfer, and projectile motion are all implicit rules in role-playing video games like Grand Theft Auto. Entertainment, argues Tom Ellman, associate professor of computer science at Vassar, is just the tip of the iceberg for physics-based animation. Science, education, and engineering are other areas in which realistic world models are important. Although building situation-specific animations is difficult and time consuming, Tom has developed software that automatically, based on inputs from an interactive human user, unfolds the physics-based world.[111] Aeronautical and automotive engineers use real-time three-dimensional animation of rigid bodies to help speed the development of new airplanes and cars. When you can build and test a vehicle on the computer, then you can bend, twist, kill, crush, and destroy it with impunity. And you can do so over and over to explore the impact of changes in the vehicle’s design.
As part of their design process using a digitally simulated world, engineers also employ what they call “genetic algorithms” (GA). The GA approach is evolutionary, using randomness to create novel variants of a design. The performance of the variants is judged using a fitness function that seeks to maximize, usually, a single aspect of performance. The variants with the best fitness are selected to mutate and mate to create a next generation of novel designs for testing.
By searching for an engineering design that optimizes a single aspect of performance, such as fuel economy, the GA process is an example of a class of procedures called “hill climbing” routines. The hill, in this case, is the specific area in your design space that combines features in a way that gives you the best possible—the optimal—results.[112] Once your evolutionary simulation finds the top of a hill, you then stop and build the winning design as a physical entity. By the way, only if that physical embodiment of the digital simulation works as predicted can your digital simulation be said to have been “validated.” In many ways, we were validating backward, having started with a physical system, Tadro3, and creating the digital simulation of it.
108
I’ve tried to address this line of criticism in more detail by placing Tadro and the Tadro evolutionary system in the explicit context of Webb’s system: J. H. Long, “Biomimetic Robotics: Building Autonomous, Physical Models to Test Biological Hypotheses,”
109
Strictly speaking, Rob’s observation is a logical inference based on the following reasoning:
110
An open-source physics engine that may be used for simulation of rigid-body dynamics is ODE: www.ode.org/. Unfortunately, this physics engine along with most others does not model the interactions of flexible bodies and fluids. That’s because the physics are very complicated. Thus, we’ve been building and modifying our own physics engine.
111
If you are interested in Tom’s approach, try: Thomas Ellman, Ryan Deak, and Jason Fotinatos, “Automated Synthesis of Numerical Programs for Simulation of Rigid Mechanical Systems in Physics-Based Animation,”
112
The traits that the engineer allows to vary, just like we allowed traits to vary in our Tadro3, defines the design space. The hill, then, is the combination of those traits that gives the best performance compared to other combinations. Engineers use genetic algorithms instead of doing exhaustive searches of all possible combinations of traits; if you have many traits or features you need to consider, using a genetic algorithm can help you find the hill faster because you don’t rely on your intuition to find the optimum (as defined by a position in multiple dimensions, e.g., the optimal design has a specific weight, drag coefficient, and gear ratio).