BIOLOGICAL INSPIRATION FOR COMPUTING 271
In this case, the goal is to evolve a neural network that has the potential to learn things, rather than
evolving the things themselves that are the object of learning. In the case of a robotic brain, it is too
difficult to anticipate all of the possibilities that might face the robot, and thus it is impossible to develop
a fitness function that fully reflects this diversity. By giving the brain the ability to learn and reason, one
can circumvent this difficulty, and as long as one can develop a fitness function for how well the brain
has learned over some period, evolutionary techniques can be used to evolve a robotic brain. (Note that
the indirect nature of this approach makes it doubly difficult to understand what is going on.)
An example of such work is that of Sims (Box 8.2).
8.3.1.7 Behavior of Evolutionary Processes,
Today, those working in evolutionary computation are not able to predict, in general, how long it
will take to evolve some desired solution or determine a priori how large an initial population size
should be, how rapidly mutations should occur, or how often genetic crossovers should take place.
Obviously, all of these parameters have some potential impact on the rate of evolution and how effec-
tive a solution might be. Yet how they should be set and their possible relationship to the nature of a
given problem are, in general, not known, although some intuitions exist in this area.
Box 8.2
Genetic Programming in AnimationIn the world of computer graphics and animation, it can be difficult to build virtual creatures that behave in a
realistic manner and simultaneously remain under the user’s direct control. For example, directly controlling
the positions and angles of moving objects such as limbs can result in detailed behavioral control, but likely
at the expense of achieving physically plausible motions. On the other hand, providing a realistic, physics-
based environment in which the relevant dynamics are simulated can result in a higher degree of realism, but
will likely make it difficult to achieve the desired behavior, especially as the entities involved become more
complex.One way to manage the complexity of control is to optimize the behavior of the creature against some fitness
function. Using evolutionary techniques, it is possible to fabricate creatures that behave realistically without
understanding the procedures or parameters used to generate them. Different fitness functions can represent
different modes of movement (e.g., swimming, walking, jumping, following a source). This approach forces
the user to sacrifice some detailed control, but there is also considerable gain in automating the creation of
complexity—and the user still influences the outcome by specifying the fitness function.For purposes of animation, a creature is determined by its physical morphology (e.g., size, shape, number of
legs) and the neural system for controlling the relevant muscle forces (the neural system involves sensors that
tell the creature about its immediate environment, effectors that cause motion [analogous to muscles], and
neurons that retain some memory of its previous states). Both morphology and neural system can be evolved,
resulting in a succession of increasingly “fit” creatures that move realistically in a given mode.In Sims’ work, a developmental process was used to generate the creatures and their control systems. The use
of such a process allowed similar components, including their local neural circuitry, to be defined once and
then replicated, instead of requiring each to be separately specified. Thus, a coded representation—a geno-
type—of a creature was established that uniquely defined the phenotype of that creature—its morphology and
neural system. By evolving the genotype, different phenotypes emerged.SOURCE: Adapted from K. Sims, “Evolving Virtual Creatures,” Computer Graphics, Annual Conference Series (SIGGRAPH ‘94 Proceed-
ings), July 1994, pp. 15-22.