As the final section of our design document, we will discuss the features we
wish we had time to add to the system, but, due to time and resource
constraints, must wait until later versions of the project.
- The original goal of this project was to implement agents
which could use actual planning strategies to plan
out moves in the long-term, instead of the simple one-step lookahead provided in this version.
Due to time constraints and the complexity of the issue, we were unable to tackle this. We expect
that in the future this will be accomplished, however, as several members of the team will continue
working on the system as part of a research project over the next year.
- It has been mentioned that the general heuristic the agents
currently use is extensible. Many more heuristics could be defined;
as many ways as one could think about how to play the game Acquire.
Since the heuristics are generally defined as the maximization of
certain parameters, one could even implement non-intuitive
strategies to determine if the best strategy for winning Acquire is
not an obvious one. These extensions to the current General
Heuristic will be implemented as a parameter passed to each Agent
that tells it which Heuristic to load.
- The system has been implemented around the Acquire domain; however, we believe the same
statistical analysis of viable strategies would be useful in many other
domains, for example, engineering
design, day trading, and so on. The system as it is implemented will
provide a clear API for future programmers who wish to implement another
domain for which the agents must
perform certain actions to reach some
goal.
- The Acquire system in its present incarnation will allow only 2-6
computer agent players. It may be interesting (or fun) for humans to
pit their own strategies, which are inherently more adaptable than
software agents, against the computer. The system will be easily
extensible to simply "plug in" human input to the engine
where it expects agent
input.
- The statistical analysis of the game(s) played by the agents utilizing
specific heuristics is currently displayed simply as a tally of the
number of games played and which agents won using which strategies, or
heuristics. We consider that users may want to view the progress of a
game, to use their intuition to recognize when an agent is playing
poorly or well. Also, it is much easier for a human player (at such
time as this is supported, see above) to follow the game if presented
with a graphical interface.
- Agents are not intended to be capable of collaborating with each other
during game play in this version of the system. In some future
incarnation, however, it could be possible that this would be allowed
and also that the user could define the level of privacy afforded to
agents who are whispering to
each other, ranging from full privacy (i.e., no one knows the pair
communicated), to medium privacy (i.e., the other agents are aware that
some communication was taking place), to no privacy (i.e., the other
agents both see and hear what was communicated).
[Go