“Evolving” Robots Challenge Evolution
“Evolvable” robots invented by a team of computer scientists from Brandeis University, Waltham, Massachusetts, headlined the popular media last summer.1 According to the press, these machines were capable of developing, on their own, new and better machines, thus mimicking, in the laboratory, the evolutionary process biologists ascribe to nature. Reporters heralded the work as a key achievement in the area of artificial life, also citing it as powerful supporting evidence for natural-process evolution.
Details of the original report reveal, however, that though this research represents an important advance in the field of artificial life, most media overstated the researchers’ accomplishment.2 On closer examination, the work on “evolvable” robots constrains, rather than supports, biological evolution. The constraints it reveals raise serious problems for the evolutionary paradigm, calling into question the likelihood that biological evolution could proceed as a creative process unattended in nature.
The Brandeis researchers developed a virtually (rather than physically) “self-evolving” robot. Specifically, they used a computer simulation to “evolve” (or create progressions of) virtual robotic designs. They pre-selected two types of building blocks, rods and linear actuators, with the aim of generating (via the computer simulator) a robot that could move itself horizontally. By randomly altering structures again and again––each time selecting the best design elements, while discarding those it determined inferior—the computer simulator produced a robot design with a capacity for horizontal movement. The computer then produced a plastic prototype of the robot, and the researchers manually attached a motor to the robot to power its movement. The experimental protocol offered no opportunity for feedback from the physical world into the computer-directed “evolutionary” process.
The Brandeis scientists thus developed, for the first time, a successful interface between virtual search-and-optimization “algorithms” (step-by-step computerized problem-solving procedures) and a technology capable of producing a physical prototype.
Genetic algorithms (used for some time by engineers and scientists to search for optimal designs, including robotic designs3) mimic the process evolutionary biologists think is responsible for driving evolutionary change. A genetic algorithm begins by evaluating a number of initial designs according to a predetermined set of criteria. The most “fit” of these designs are modified by simple changes (called mutations, or asexual reproduction) or by combining parts of two or more designs (called sexual reproduction), and the new designs are then evaluated. The genetic algorithm repeats this process over and over again until a superior design (one that optimally fits the selection criteria) emerges.
Using genetic algorithms, the Brandeis scientists sought a robotic design capable of horizontal movement. The genetic algorithm started with two hundred random designs, with rods and linear actuators as the predetermined building blocks. A design that yielded a certain amount of movement in a certain amount of time would be “selected.” The most successful designs were used to produce a new set of designs by randomly modifying, removing, adding, or relocating rods and linear actuators. Typically, tens of generations passed before machine designs capable of any movement “evolved.” The “evolution” of working designs required 300 to 600 generations and about 100,000 designs.
The “evolutionary” process used by the Brandeis researchers produced a variety of fairly sophisticated machines. The robots moved by dragging, ratcheting, crawling, scooting (crab-like sideways motion), and side-to-side oscillations, to name a few. Robots typically employed about 20 building blocks (rods and actuators) in various configurations. These robust robotic designs retained the capability for movement even after researchers removed or altered the size of various building blocks.
Though the work of the Brandeis scientists might seem to lend empirical support to biological evolution, careful evaluation of this work uncovers several significant constraints that make biological evolution unworkable in nature.
Selecting Building Blocks
The choice of building blocks is critical to the evolutionary process. In this case, the Brandeis scientists thoughtfully selected rods and linear actuators. The simplicity of these building blocks allowed for maximal architectural and manufacturing flexibility. Thus, the intelligent designers provided the genetic algorithm with optimal features, giving ample opportunity to find a workable design. Selection of the wrong building blocks would have limited the design options to the point that few, if any, workable designs would be possible.
Nature offers a variety of building blocks, not all necessarily optimal. And, according to the evolutionary paradigm, nature offers no “intelligence” or process to select the appropriate building blocks to ensure the availability of the largest number of workable design options.
Modifying Failed Designs
The computer simulator was able to find workable robot designs only after many generations had passed. Tens of generations were required before any of the robotic designs acquired the capability for movement. The genetic algorithm kept the “evolutionary” process going, but in nature, if the selection criteria are not met, the evolutionary process comes to an abrupt halt. Inability to meet selection criteria means death of the organism and, more importantly, the failure to propagate the next generation. Biological evolution cannot modify failed or dead organisms to find one that will survive and yield the next generation.
Evaluating Transitional Forms
The relatively simple selection criterion (horizontal movement) required the computer simulator to evaluate about 100,000 designs. Selection criteria for a biological organism are far more complex and multifarious; many more designs would have to be explored randomly before a workable design for a living organism could emerge.
If the evolutionary process were at work in nature, multitudinous transitional designs would connect various organisms. Therefore, a large number of transitional creatures should appear in the fossil record. The near absence of transitional forms, one of the hallmark features of the fossil record, 4 stands in sharp contrast to the expectations of evolution.
Efficient search among the robotic design possibilities appears to be dependent upon the dramatic rearrangement of designs with each generation. The algorithm employed by the Brandeis researchers modified, deleted, added, and relocated the components of successful designs after each round of evaluations. In biological systems, no mechanism exists to produce these dramatic changes. Rather, the mechanism that produces biological change (mutations) can only yield small variations on existing biological features.
The Brandeis scientists set out to “evolve” movable robots apart from any human intervention. However, human influence, and hence, intelligent design, permeated the entire experimental setup: the development and application of the genetic algorithm, the choice of building blocks and selection criteria, the manual addition of motors, and the provision of materials for production of physical prototypes. As one engineer from the Massachusetts Institute of Technology (MIT) commented, “The resulting machines cannot match the complexity of the rapid-prototyping machine designed by human engineers that is required to do the actual fabrication.”5
The research program in artificial life will help researchers discover boundary conditions for natural biological processes. The results of this Brandeis study lead some scientists to anticipate that additional work will further challenge evolutionary theory and support the case for intelligent design.
- For example, “All Things Considered,” with Robert Siegel and Linda Wertheimer, National Public Radio, August 31, 2000.
- Hod Lipson and Jordan B. Pollack, “Automatic Design and Manufacture of Robotic Lifeforms,” Nature 406 (2000), 974-78.
- Rodney Brooks, “From Robot Dreams to Reality,” Nature 406 (2000), 945-46.
- Niles Eldredge, The Pattern of Evolution (New York: W. H. Freeman, 1999), 133-45.
- Brooks, 945-46.