Information Technology Reference
In-Depth Information
The question for research on computational creativity then becomes how can we
develop CAD-like interactive environments for enabling experimentation for other
creative tasks? Thus, our research laboratory has been studying the nature of exper-
imentation in creative tasks such as technological invention and scientific discovery,
including thought experiments, simulated experiments, and real experiments. For
example, we have studied how scientists construct explanatory models of complex
phenomena through thought experimentation [
20
,
34
,
35
]. We have also developed
interactive tools for supporting simulated experiments in both task domains. For
example, we have developed interactive tools that enable science students to con-
struct conceptual models of complex phenomena, evaluate the models through sim-
ulation, and revise the models based on the results of the simulation, repeating the
cycle until the student has constructed an acceptable model [
33
,
39
,
76
].
In this article, we view another creative task domain through this lens of experi-
mentation as a central element of the creative process: design of game-playing intel-
ligent agents. Human often play interactive games against one or more autonomous
game-playing software agents. In general, most humans are more likely to enjoy
playing interactive games against intelligent software agents instead of trivial agents.
Thus, an important question in design of interactive games is how to design intelli-
gent game-playing agents. This is a difficult task: in its most general form, it appears
equivalent to the general task of designing AI agents! The complexity and creativity
of the task depends on the nature and scale of the game. In this article, we focus on
turn-based strategy games such as Freeciv.
1
Multi-player turn-based strategy games
such as Freeciv are dynamic, only partially observable, non-deterministic, and often
have huge state spaces. Thus, design of software agents that can play such turn-based
strategy games is a very complex and creative task.
In this article, we describe a CAD-like interactive environment for designing
game-playing software agents called GAIA that enables experimentation with an
agent's design. GAIA provides the designer with a visual editor for constructing a
conceptual teleological model of the agent in terms of its tasks, methods and knowl-
edge (TMK) in a high-level agent modeling language called TMKL2. GAIA also
provides the designer the ability to simulate the agent: GAIA automatically trans-
lates the TMK model into executable code, interacts with game server, and runs the
simulation. The designer may use GAIA to not only monitor the results of the simula-
tion, but also to inspect the agent's knowledge and reasoning during the simulation.
The designer may then modify the teleological model of agent; GAIA provides a
persistence mechanism for managing agent models. The designer may repeat this
cycle of experimental modeling and simulation until the designer is satisfied with
the simulation results.
While the above process of experimentation in the design of game-playing soft-
ware agents is parallel to the experimentation processes in CAD, it is important
to note that there are also important differences between the tasks of engineering
design and agent design. A fundamental difference between the two is that unlike
most engineering systems, an intelligent software agent can
learn
fromits experiences
Search WWH ::
Custom Search