Interactive Meta-Reasoning: Towards a CAD-Like Environment for Designing Game-Playing Agents - Computational Creativity Research: Towards Creative Machines - page 354

17.5.1 Agent Modeling
An intelligent agent is an autonomous entity that maps a history of percepts in an
external environment into an action on the environment to achieve goals and maxi-
mize utility [ 66 ]. Thus, a game-playing agent is intelligent if it autonomously selects
actions that accomplish its goals and maximize its utility. TMKL2 is a language
for modeling agents for the purposes of meta-reasoning, and, in particular, meta-
reasoning for agent self-adaptation.
TMKL2 expands and extends the original TMKL language [ 53 , 54 ]. TMKL is
comparable to theHierarchical TaskNetwork language (HTN) [ 24 , 56 ] for automated
planning. Planning problems in HTN are specified in terms of different kinds of
tasks including goal tasks, primitive tasks corresponding to actions in the world, and
compound tasks that are composed of simpler tasks. Constraints among the tasks
are expressed in the form of networks. Both HTN and TMKL languages emphasize
the connection between goals and methods for accomplishing them, with methods
in TMKL composing the primitive tasks for achieving higher-level tasks much like
compound tasks do in HTN. While the HTN language arose out of AI research
on planning, the origins of TMKL2 are rooted in AI research on knowledge-based
systems [ 10 ], Chandrasekaran, Johnson and Smith [ 11 ] and functional modeling of
physical systems [ 73 ]. While HTN was designed for automated planning, a very
common task in AI, TMKL was designed for meta-reasoning. Hoang, Lee-Urban
and Munoz-Avila [ 36 ] have compared the expressivity of HTN and TMKL. They
found that in principle TMKL and HTN [ 56 ] have the same expressive power, but
that TMKL is more explicit in its some of its constructs than HTN. Molineux and
Aha [ 50 ] have used a variation of TMKL for modeling game-playing agents in their
TIELT game-playing agent benchmarking system.
TMKL2 is also comparable to the Procedural Reasoning System (PRS) language
Georgeff and Lansky, Rao and Georgeff [ 29 , 55 , 63 ]. Both languages emphasize
the connection between goals and the procedures for accomplishing them, and both
support reasoning about the procedures rather than their construction on the fly, as
would be done by a planner. While PRS supports asynchronous communication with
agents in continuous, real-time settings and reactive control, TMKL2 is discrete and
synchronous. Although both languages supports alternative means for accomplishing
goals, the operational semantics of TMKL2 does not currently provide support for
dynamic choice among them. Although both languages in principle support meta-
level reasoning, later versions of PRS drop this feature for reasons of performance.
In contrast, as we noted above, TMKL2 is specifically designed to support reflection.
Finally, both languages have a database of world knowledge, although there is no
mention of PRS supporting inheritance, which TMKL2 does.
While HTN supports planning, it does not support plan adaptation. In contrast, if a
planned action taken by an agent results in a failure due to changes in the world, then
PRS can modify the action, thus making the planning reactive to changes the world.
The goals of GAIA are different. While PRS adapts a plan in the presence of a plan
failure, we seek to adapt the planner itself. Thus, while the PRS language supports