Interactive Meta-Reasoning: Towards a CAD-Like Environment for Designing Game-Playing Agents - Computational Creativity Research: Towards Creative Machines - page 340

is found (if one exists). Thus, the Generate and Test method is iterative. The balance
between generation and testing in the method depends on the knowledge available
to the agent. At one end of the knowledge spectrum, if the agent has complete
and correct knowledge of the problem and the problem world, then the agent may
generate a provably optimal solution to the problem and little testing is needed. At
the other extreme, if the agent has little knowledge of the problem and the problem
world, it may generate a candidate solution randomly, and then evaluate the solution,
for example, by applying an objective function to the solution or by trying out the
solution in the world, with the world acting as its own “objective function”. The logic
school of AI is closer to the former end of the knowledge spectrum; the evolutionary
computing school is closer to the latter.
If Generate and Test is a basic process of intelligence, it is even more fundamen-
tal to creativity. Many problems that intelligent agents encounter routinely are well
defined. For example, in planning, the agent may have knowledge about the initial
state and the goal state, as well as the actions available in the world; Further, the
agent already may have solved the problem, or similar problems, in past. Thus, a
robot planning a navigation route to drive me frommy home to my office may require
some intelligence, but, under normal conditions, not much creativity. Most problems
requiring creativity, on the other hand, are ill defined and open ended. From under-
standing a story to comprehending a movie, from painting a portrait to composing
a symphony, from inventing a technological system to discovering a scientific phe-
nomenon, for creative tasks, the agent typically has neither sufficient knowledge of
the problem and the problem world to generate an acceptable solution initially, nor
an objective function to efficiently evaluate a candidate solution. Thus, for creative
tasks, the agent must experiment: it must generate candidate solutions, evaluate the
solutions, use the knowledge gained from the testing to generate better solutions,
and repeat the cycle until it has acquired sufficient knowledge of the problem and
the problem world that it can generate an acceptable solution. Again evaluation may
take several forms, but the central element in the creative process is experimentation.
Given that experimentation is a central element of the creative process , a core
question for research on computational creativity is how can we support humans in
experimentation? Note that the design of the experiments must be such that each
experimental run produces new information about the problem or the problem world
so that the agent can try to produce a better candidate solution in the next iteration of
Generate and Test. To answer this question, we take inspiration from the enormous
success of computer-aided design (CAD) as a scientific and technological enterprise
[ 9 , 77 ]. Design in general is a creative task because novel design problems typically
are open-ended and ill defined. CAD environments are successful in part because
they enable designers to experiment with candidate solutions to design problems.
When a designer specifies a candidate solution to a problem, CAD tools can help
the designer, for example, in constructing geometric models and performing numer-
ical simulations for evaluating the solution. The geometric modeling and numerical
simulations produce new information that the designer can take into account in refin-
ing or revising the candidate solution.