Game Development Reference
In-Depth Information
sing at the funeral? The pattern-matching problem goes from the general
problem of ''Do I sing now?'' to the more subtle and finer-grained problem of
'' What do I sing now?'' If Horatio has a hybrid AI, the general AI has correctly
figured out that it is time to sing, and it is handing the situation off to a spe-
cialized book of moves AI for song selection. Since he is a tenor and not a
baritone, he is not likely to belt out the difficult, well-known line, ''Figaro!
Figaro! Figaro!'' But just the same, as an opera singer, his book of moves is deeply
loaded with equally stunning but equally inappropriate songs. The right selection
might not be Rossini or even Mozart, despite the awesome quality of their songs,
but something well known from a church hymnal. The pattern-matching pro-
blems we examined in Chapter 4, ''Rule-Based Systems,'' are still with us.
Hybrid AI
Here, the term hybrid AI means a combination of more than one kind of AI so
that the different forms mitigate the weak areas of the others. Coaches might call
plays, but players execute them. The players react in real time, adjust and make
changes, and do their best to exploit the unexpected. A book of moves by itself is
not commonly used as a complete AI, although the line blurs in rule-based
systems composed of both general rules and highly specific ones. One of the
particular strengths of a book is the ability to recognize the value or peril of a
situation that a more general system overlooks. We will see this in various
applications.
Chess
Chess is well suited to a hybrid approach. The Deep Blue Chess computer
combined powerful search capabilities with an opening book of 4,000 positions,
an extended book drawn from 700,000 grandmaster games, and a database of
endgames [Campbell02]. In 1997, this software, running on massively parallel
hardware that included custom Chess chips, was the first Chess program to beat a
reigning world champion Chess player. The evaluation function of the search was
astoundingly rich, but the various topics helped detect situations the search
would rank improperly or spend too much time evaluating.
While interesting as a thought problem, Chess is hardly suited for programmers
just starting to write AI. Games with simpler rules might appear more
approachable, but it depends on the game. Go is harder for machines than Chess,
but steady improvements suggest that in 20 years, machines may be able to
 
 
 
Search Nedrilad ::




Custom Search