Interactive Meta-Reasoning: Towards a CAD-Like Environment for Designing Game-Playing Agents - Computational Creativity Research: Towards Creative Machines - page 358

Finke, Ward and Smith [ 25 ] proposed a model of ideation called Genexplore in
which “preinventive structures” are first generated and then explored. Importantly,
the processes of generation and exploration in Genexplore are subconscious and thus
unavailable for introspection. Benami and Jin [ 5 ] have operationalized this model to
ideation in conceptual design. In contrast, we posit experimentation as a deliberate
methodology for creative tasks such that the results of the experiment are available for
inspection and reflection. Cook, Colton and Gow [ 15 ] view inspection, modification
and execution of code of video games as an example of computational creativity.
Their work is similar to ours in that both emphasize the importance of evaluation in
the design process and in both the program code is generated automatically. However,
we view experimentation, not merely evaluation, as a central element of creativity.
Secondly, we postulate meta-reasoning as an important cognitive process of cre-
ativity. Meta-reasoning allows the agent tomonitor not only its behaviors in the world
but also its knowledge and reasoning in producing the observed behaviors. Meta-
reasoning enables the agent to reflect and introspect on its behaviors, knowledge and
reasoning, and to revise its knowledge and reasoning to modify its behaviors. This
process of observation, reflection, and adaptation appears to be a cognitive process
of creativity.
17.6 Summary, Conclusions and Future Work
We posited that because most creative tasks, such as design, invention and discov-
ery, tend to be ill defined and open-ended, the creative process is characterized by
experimentation. The experimentation in creative tasks typically entails the genera-
tion of hypotheses and models, and the evaluation of the hypotheses/models. Thus,
we posited that a fundamental question for computational creativity is how can we
support experimentation.
We investigated this question in the context of the creative task of designing
software agents for playing the turn-based strategy game of Freeciv. We described
GAIA, a CAD-like interactive environment for designing software agents for playing
Freeciv. GAIA provides a visual editor for constructing Task-Method-Knowledge
models of an agent in the TMKL2 language. GAIA has the capability of translating
the TMK model of the agent's design in TMKL2 into program code and executing
the agent in the Freeciv world. The designer may observe the behavior of the agent
and monitor its reasoning and knowledge states. If the designer is not satisfied with
the agent's behavior, he can revise the TMKmodel of the agent, and repeat the above
process until he finds an acceptable agent design. In this way, GAIA enables the
designer to experiment with the design of the game-playing agent.
We explored the role of meta-reasoning in the creative process. Given an agent
endowed with a model of its own reasoning and knowledge, the agent may use its
self-model tomonitor its reasoning and knowledge states of the agent in addition to its
behaviors and interactions in theworld. Given a failure in the agent's interaction of the
world, model-based meta-reasoning also enables generation of hypotheses about the