Testing
The Staykov Strategy Finder was extensively play tested against humans
in the games DPPA, DPPB, etc., and at the London Computer Olympiad.
Gilles Schaeffer's improvements have been "debugged," and
the retreats now work according to the new system. It has been tested
with pretty complicated examples (high numbers of interacting
retreats). It is better and significantly faster than the
previous version.
However,
the various improvements have not been tested via self-play, nor
against humans. It would also be appropriate to test the observation module at the same time.
Winter
In the next issue of The Diplomatic Pouch, we will discuss the algorithms
used to search for good builds and removals during an adjustment
phase.
However, more work is need to make builds compatible with retreats in
the sense
that the program must be modified to consider the possibility of removing
a retreating unit and building it elsewhere.
Ideas:
- Do Fall and Winter as normal, but while doing the seconfnd phase of build
optimization, modify retreats as well. However, if the SC count changes, then
what do we do!?
- Solve each set of corrolated retreats independently, given
different numbers of disbands. Then for each country and for each
possible number of disbands, pick the best combination of builds and
retreats. Compare. Then optimize, how?
- If #Ply=2 or more, then count "estimated SC holdings" assuming
everyone grabs SC's during retreats. If a country's home SC's are
threatened (distance < 3, say, to enemy), and if
the expected number of SC's is not less than the number of units for a country,
and that is less than the number of open home SC's, then for each unit
of that country, add all possible builds to the list of retreats. Then
compute sets of corrolated retreats and optimize as before....
- ????
Return to main article