You are on page 1of 8

AI Magazine Volume 9 Number 3 (1988) (© AAAI)

Minimaxing
Theory and Practice
Hermann Kaindl

Empirical evidence suggests that searching


deeper in game trees using the minimax
E xisting models for analyzing
properties of minimaxing seem
to have been designed primarily to
here is a bounded look-ahead to
artificial terminal nodes.
Except in very rare cases such as the
propagation rule usually improves the support formal analysis, rather than game nim, there is no known way of
quality of decisions significantly. Howev- the proper modeling of the games for directly evaluating the exact status
er, despite many recent theoretical analy- which minimaxing has proved very (the true value) of such artificial ter-
ses of the effects of minimax look-ahead,
successful in practice. The discovery minal nodes. Consequently the use of
however, this phenomenon has still not
been convincingly explained. Instead,
and analysis of pathological game heuristic estimates is usually the only
much attention has been given to so- trees, for which minimaxing does not practical resort. Such heuristic values
called pathological behavior, which occurs work, is important but seems to have are assigned by so-called static evalua-
under certain assumptions. This article no correspondence to practical obser- tion functions. In practical applica-
supports the view that pathology is a vations. Conjectures about practice tions these heuristic values are usually
direct result of these underlying theoreti- based on models with assumptions not restricted to the game-theoretic
cal assumptions. Pathology does not occur that do not apply are rather dangerous. values, but range over an interval of
in practice, because these assumptions do This article takes the opposite integer values. Theoretically, real
not apply in realistic domains. The article approach to understanding the
presents several arguments in favor of
numbers can also be used. While there
benefits of minimaxing. After summa- is no satisfactory theory of the actual
minimaxing and focuses attention on the
gap between their analytical formulation
rizing existing models according to semantics of these numbers, it is intu-
and their practical meaning. A new model their practical relevance, we present a itively clear that they should induce a
is presented based on the strict separation new model that was developed primar- partial order in the various positions
of static and dynamic aspects in practical ily from observations. This model of a game according to their worth in
programs. finally, certain methods of seems to capture the essential features the sense of “goodness” or “strength”
improving minimax look-ahead are dis- of the method, but introduces new for one side. Conceptually it is rather
cussed, drawing on insights gained from complications, despite efforts to sim- problematic to map all considerations
this research. plify it as much as possible. We con- about a position into a single number,
clude with a discussion of the conse- especially as there is no measure of its
quences of such investigations for the reliability. Nevertheless, up to now
proper use of minimaxing. this simple scheme has been superior
to more complicated ones in game-
Background playing practice.
Under the assumption that the look-
First, let us briefly review the concept ahead tree (graph) of moves for both
of minimaxing. Most current comput- sides is searched deeper than one ply
er programs for two-person, perfect- (the technical term for a half-move, or
information games (including chess, move by one side), the natural ques-
checkers, and kalah) use minimaxing tion arises as to what to do with the
as their basic method for choosing a heuristic values of the terminal nodes.
move from a given position. While More precisely, how can they be used
such games could be completely for making a reasonable move deci-
solved in principle using game theory, sion? When exactly one ply is
the combinatorial explosion inherent searched, it seems clear that the best
in “interesting” games makes this choice is the move resulting in the
totally infeasible within any practical successor position with the best value
time and space limits. Hence, in con- (the maximal one, assuming that high-
trast to the usual attempt in one-per- er values are better for the side that is
son games (problems) to search for a moving). When there are several such
real goal-node, the usual paradigm moves, one of them may be chosen

FALL 1988 69
rience shows that decisions based on
this method are, with only very few
+3 exceptions, much worse than those
MAX
based on deeper and deeper searches
using minimaxing.
While the basic concepts of mini-
maxing (with the exception of search
techniques and backward pruning)
were proposed very early by two
famous researchers, Shannon (1950)
+3 -1 MIN
and Turing (1953), the success of this
technique remains even today a theo-
retical mystery.

Is Minimaxing Pathological?
Nau (1979, 1980, 1983b) has proved
+3 +9 -1 +4 MAX that for certain classes of game trees
the decision quality is degraded by
searching deeper and backing up ter-
minal values using the minimax prop-
agation rule. Nau called such behavior
pathological. The key assumptions for
these game trees follow.
+3 -2 +4 +9 -2 -1 +4 -5 1. The trees are uniform, in that each
nonterminal node has exactly m + n
children nodes.
2. Every “critical” node has m chil-
Figure 1. An Example of Minimaxing. dren with the same true value, fol-
lowed by n children with the opposite
arbitrarily. But how should the heuris- ignore large parts of the look-ahead true value (for the remaining nodes,
tic estimates be backed-up (propagat- tree without changing the result. the true values of the children are
ed) from deeper in the tree? There are algorithms that guarantee identical according to the game-theo-
For a long time, there was universal the same minimax value and move retic relationship).
agreement to proceed as follows: If one choice while saving considerable 3. There is an independent (identical)
player (called MAX) is on move from a effort. The best-known of these is the distribution of error by the static eval-
position within the tree, take the max- alpha-beta algorithm which, for uation function that estimates values
imum of the successor values; if the instance, in figure 1 must only search for terminal nodes.
other player (called MIN) is on move, the part of the tree shown in boldface. Recently, Schrüfer (coauthor of a
take the minimum. By use of this Since these backward-pruning possi- competition chess program) modified
back-up rule, a so-called minimax bilities are not of primary interest in this model, distinguishing two differ-
value of the whole tree can be com- this article, the reader is referred to ent error parameters for “overestimat-
puted recursively. figure 1 illustrates Campbell and Marsland (1983) or vari- ing” and “underestimating,” respec-
the way values are propagated toward ous textbooks on AI. tively (Schrüfer 1986, Section 1).
the root. Usually, the primary interest An important question, then, is how While his finding that pathology can
is not the minimax value of the tree deeply the different branches should be avoided—given low error rates—is
itself, but rather the move to be select- be searched within the given time lim- quite illustrative, the requirement for
ed. In accordance with the back-up its. In practice, there is a tendency to a negligible probability of just one of
rule, the arc (move) leading to the best search each branch to about the same the two errors (underestimating) could
(by convention the maximal) backed- depth (of course, with the exception of not be related satisfactorily to observa-
up value of the immediate successors backward pruning), although there is tions in computer chess practice.
is chosen. More precisely, the choice an important potential for improve- In later work, Nau (1982, 1983a,
is one of what may be several moves ments through searching to variable 1983c) investigated a class of “real”
leading to the maximum. depth (Kaindl 1983b). However, from a games, called Pearl’s games or P-
While any reasonable method of theoretical point of view the crucial games, in which pathology actually
searching can be used, the usual question is: Why search at all? As occurs. Similar classes of games, called
choice is an iteration of successively heuristic values have to be used any- N-games and G-games, were shown
deeper depth-first searches in the form way, those of the immediate successor not to be pathological. What is the
of backtracking, to minimize storage positions may be used directly for essential difference between P-games
requirements. In fact, it is possible to selecting a move. Unfortunately, expe- and N- or G-games?

70 AI MAGAZINE
Our description of such games can WIN
be rather cursory, since they are amply
9 W L L W
described in Nau (1982, 1983a, 1983c).
A P-game is played on a board measur- L W L W
ing bk/2 by bk/2, where b > 1 and k > 0 W W W L
are integers. For the example shown in
L L W W
figure 2, which is taken from Nau
(1982, 1983a) b = 2 and k = 4 were cho-
sen. The initial configuration of the
board for a P-game is constructed by WIN WIN
randomly assigning each square of its *
board one of two possible values, inde- 4 W L 5 L W
pendently of the values of the other L W L W
squares. We have indicated these val- MAX
W W W L
ues as W and L to emphasize their cor-
L L W W
respondence to the true values WIN
and LOSS on leaf nodes. The first play-
er moves by dividing the board verti- WIN WIN LOSS WIN
cally into b sections of equal width, 2 W L 2 W W 2 L W 3 W L MIN
and discarding all but one of the sec-
tions. The opponent’s move consists L W L L L W W W
of doing the same horizontally with
the remaining part of the board. The WIN WIN WIN WIN WIN WIN WIN
LOSS
play continues in this manner until
only one square is left. If the square 1 W 1 L 1 W 1 W 0 L 2 W 2 W 1 L
MAX
has value W, the last player (MAX) L W L L L W W W
wins. If it has the value L, MAX loses,
and the opponent (MIN) wins.
W L L W W L W L L L W W W W L W
Figure 2 shows a complete game tree
for this example, with the true values
WIN and LOSS indicated for the Figure 2. A Game Tree for a P-Game with b = 2 and k = 4.
boards from MAX’s viewpoint. The
numbers denote the heuristic values
of these boards, computed by the stat- strength or weakness of a board posi- tion function (when evaluating termi-
ic evaluation function described in tion tends to be roughly the same for nal nodes) is a necessary condition for
Nau (1982) (by simply counting the sibling nodes. This dependence of val- pathology in minimaxing.
number of W squares). The interested ues is sufficient to prevent the occur- Michon (1983) found nonpathologi-
reader is encouraged to figure out rence of pathology. Nau’s investigation cal behavior for games with “inert”
whether a 1- and a 2-ply search using of G-games (Nau 1983c) also shows structure, despite the assumption of
minimaxing will choose the correct nonpathological behavior. These independent terminal values. Conse-
move for MAX in the position marked games are a modification of P-games, quently, pathological behavior of mini-
by an asterisk. Note that for k[<=] 7, such that sibling nodes have many maxing seems to be restricted to
even for P-games, no pathological children in common, which results in games with the properties of indepen-
behavior has been found in Nau a different type of correlation between dence of nodes as well as of noninert-
(1982). them. ness (games with constant branching
Due to the board-splitting character Essentially the same findings were degree are noninert).
of these games, the values of the real reported independently by Beal (1980, However, Pearl (1983, p. 452) still
leaf nodes directly correspond to the 1982) and Bratko and Gams (1982). left room for doubt by stating:
(randomly assigned) values of the ini- They began by assuming independent “Error amplification due to mini-
tial board configuration. This causes distributions for the true as well as the maxing is an established fact
the values of sibling nodes to be com- heuristic values, and then rejected one which may significantly degrade
pletely independent of each other, of these assumptions in favor of the quality of decisions in practi-
which results in the occurrence of dependence as a “clustering” of true cal games. The absence of search-
pathological behavior of minimaxing values. A model in Schrüfer (1986, depth pathology in common
in P-games. Section 2) also relates the effects of games only means that the deteri-
N-games have rules identical to deeper minimaxing more generally to oration in decision quality due to
those of P-games, with the one essen- the distribution of true values. Thus, minimaxing is masked by other
tial exception that the initial playing it seems clear that independence of processes and so is not sufficient
board is set up differently, in an the true node values or independence to show up as a weakening ability
“incremental” manner, so that the of the errors made by the static evalua- with increasing search-depth.”

FALL 1988 71
Rate P3 P4 P5 P6 P7 P8
P3 1091 4
P4 1332 16 5 1
2

P5 1500 14 1
2 4 1
2

P6 1714 15 1
2 2 1
2

P7 2052 17 1
2 3 1
2

P8 2320 16 1
2

Table 1. Tournament of the Chess Program BELLE.


Version Pi searches to a depth of i plies and then enters its quiescence evaluation; for example, P8 scores 16 1/2 : 3 1/2 versus P7.

This conclusion seems to be based ships among pieces on the board are Why Is Minimaxing
on the following statement about strong, and they always change incre-
chess: mentally when the pieces are moved.
Beneficial in Practice?
“The values of successor positions The values of positions as well as the Thompson (1982) conducted an experi-
appear tied down to the value of errors of evaluation depend on these ment on the influence of search depth
the father position partly because relationships. If one player in such a on the playing strength of full-width
we only consider normally the game had twice as much material as searching chess programs. Table 1
select set of successors which are the other, it would be absurd to (from Thompson 1982, p. 56) shows
reasonable to play. It is hard to believe that the values of descendant the results from this “BELLE tourna-
conceive of a chess position so nodes would be randomly distributed. ment.” Thompson found that an addi-
strong that it cannot be spoiled Furthermore, since all the nodes visit- tional ply of search with an identical
abruptly if one really tries to ed during a search are connected by a static evaluation function is equiva-
make a stupid move” (Pearl 1983, graph, the search tree, it is unrealistic lent to about 250 Elo rating points (Elo
p. 444). to assume their independence. Conse- 1978). Although such a formula seems
On the contrary, it is not even nec- quently, the notion of pathology in to be a doubtful means for extrapolat-
essary to conceive of such positions, such a domain is also unrealistic. ing the speed necessary for a machine
but only to look into those actually The argument given here is not to defeat the human World Champion,
investigated by the full-width search intended to belittle the theoretical it is consistent with the experience
of a real chess program. Since such a work on minimax pathology. Of gained from computer chess tourna-
search tries stupid moves most of the course, the discovery that minimaxing ments. A similarly dramatic benefit of
time, most of the terminal positions can be detrimental under certain con- searching deeper using the minimax
are so strong for one side that they ditions and the investigations of these propagation rule seems to be true for
cannot be spoiled by a single stupid conditions are important. However, checkers and kalah.
move. See also Beal (1983, p. 167). Of these findings should not be interpret- Although demonstrating this benefit
course this clustering of values in ed so that readers are tempted to appears to have been the initial moti-
games like chess is hard to prove for- believe that minimaxing is in princi- vation for the theoretical analysis of
mally, but it should be noted that Beal ple a bad method. Practical programs minimaxing, the discovery of patho-
(1982) examined the king and pawn for two-person games based on mini- logical behavior in minimaxing on cer-
versus king ending (as a complete maxing give good results, and clearly tain unrealistic game trees was the
database on this ending is available) better ones than programs using the first surprising result. In addition, the
and concluded that the clustering fac- other methods tried so far. Therefore, succeeding investigations could not
tor suffices to make minimax look- isn’t it at least as important to investi- provide theoretical support for the dra-
ahead beneficial and not pathological. gate more thoroughly those conditions matic benefit observed in practice. For
Regardless of whether such cluster- that lead to the beneficial behavior instance, Table 5 in Nau (1982) shows
ing really occurs in practice, it should observed in practice? As these condi- only modest improvements in deci-
be clear that neither the true values of tions are not understood well enough sion quality with increasing search
nodes nor the errors of static evalua- for formal treatment, one should at depth for N-games.
tion are independent in game trees of least try to approach such an under- The dependencies in real game trees
chess and related games. The relation- standing by intuitive reasoning.

72 AI MAGAZINE
reflect the structure of real games, and
consequently pathology does not
occur there; but clustering relation- 8 8 8
ships between the true values alone
cannot account for the large benefits 7 7 7
observed, mainly because they seem
to be an outer phenomenon. This rela- 6 6 6
tionship of the true values may be
induced at least partially by a more 5 5 5
fundamental relationship between
true and heuristic values within whole 4 4 4
subtrees.
Let us briefly look at some extreme- 3 3 3
ly simple examples from the very
domain for which Shannon (1950) and 2 2 2
Turing (1953) made their proposals
(and where minimaxing is truly use- 1 1 1
ful). figure 3 presents fragments of
chess positions showing simple tacti- f g h f g h f g h
cal patterns. These examples will give
(a) (b) (c)
the reader a very crude idea of what
happens in a typical computer chess
program. To follow these examples, it
Figure 3. Fragments of Chess Positions Showing Tactical Patterns.
is only necessary to know how chess
pieces move, and that the relative
heuristic values of queen, rook, and interested reader is encouraged to that a very strong improvement is nec-
pawn are about 9, 5, and 1. figure out how a minimax search may essary to combat pathology, but this is
In the following, we assume a static discover whether a rook can be gained based on the artificial assumption of
evaluation function that simply sums by the “forking” move of the White the independence of values. Thus, a
up the relative heuristic values of the pawn from g2 to g4, depending on the similar analysis based on dependent
pieces on the board for each side and remainder of the board partly shown values might be of interest.
then subtracts these sums. (Although by figure 3b. (As a hint, possibilities For the game of chess, no such
the evaluation functions actually used such as one of the rooks putting the improvement in accuracy can be seen
in chess programs may include various White king in check may be crucial.) when considering the static evaluation
Nevertheless, such issues usually functions used in actual programs.
positional and strategic subtleties, the
occur in more complicated combina- However, to compute a minimax
material term is by far the dominating
tions. For example, in figure 3c White value, the relation of the terminal val-
one.) In figure 3a a full-width, 1-ply
has to exchange queens before the ues to the minimax value is impor-
search will select Rh3xQh7 (rook
forking pawn move can be successful. tant. This is utilized by the TEST pro-
moves to square h7 to capture queen)
These examples have only shown cedure of the SCOUT algorithm (Pearl
for White on move. Since this move
rather simple tactical issues. Certain 1980). As most of the terminal values
changes the static values drastically, a
positional issues, such as how to of a chess tree are very strong for one
position containing this pattern can-
maneuver a piece to a central square, side (as was argued before in favor of
not be considered quiescent. (The are handled in a similar way. Actually, clustering), these relations are general-
importance of quiescence had already there is a strong temptation to think ly evaluated very accurately, and this
been recognized by Shannon and Tur- that the reason for the benefits of relative accuracy improves with
ing, and this concept will be treated searching deeper is trivial. But do any increased search depth. See also Bratko
more theoretically in the next sec- of the previously published attempts and Gams (1982, p. 12-14).
tion.) In fact, such simple issues are to prove the benefits of minimaxing Another argument investigated by
successfully handled by most pro- model convincingly what these issues Pearl (1983) is that of “the avoidance
grams within their capture quiescence really signify in an abstract minimax of traps.” Again the model is oversim-
search: selectively searching captures tree? plified, considering only actual termi-
even beyond the horizon of a full- A common argument is that of nal nodes (that is, mates and stale-
width search. However, when the side “improved visibility,” which was sim- mates) at all levels of the game tree as
on move has an en prise piece—a piece ply interpreted and modeled by Pearl “traps.” Pearl’s analysis shows that
in risk of capture—more subtle issues (1983) in the sense that the accuracy of the presence of such traps makes min-
become relevant (Kaindl 1982, 1983a). the static evaluation function imax look-ahead beneficial, mainly
Especially, multiple en prise pieces improves as the game proceeds. Math- because such real terminal nodes can
raise more difficult problems. The ematical analysis of this model shows be evaluated without error. Again this

FALL 1988 73
model is based on the assumption of nately, the dominating factor for eval- A Model Based on Quiescence
independence of all the other values, uation in chess, checkers (material
which probably explains why the exis- balance), and kalah (number of stones) The most commonly used technique
tence of real terminal nodes has such a is sufficiently reliable most of the to counter the horizon effect is the
strong influence on the analysis. time. It is also interesting to reconsid- quiescence search. The importance of
In computer chess practice this type er the difference between P- and N- quiescence has been known ever since
of trap alone does not seem to make a games from this point of view. Accord- the early work on computer game-
strong contribution to the benefits of ing to the tables in Nau (1982, 1983a), playing; Turing (1953) called quiescent
minimaxing. Statistics compiled at the same static evaluation function is positions “dead.” However, there is an
the fourth World Computer Chess clearly more accurate for N-games important point often missed in this
Championship with the program than for P-games. (D’ (1, k) provides a regard: Quiescence is not only related
MERLIN suggest that the number of measure of the accuracy of the evalua- to the properties of a certain game, but
actual terminal nodes in a search tree tion function for nodes with distance even more so to the static evaluation
function used. A position is nonquies-
cent if its value can be changed drasti-
cally by moves or move sequences. For
For the game of chess, no such improvement in example, in chess the usual capture
quiescence search has to be considered
accuracy can be seen when considering the static in relation to the common use of
evaluation functions used in actual programs. materially dominated static evaluation
functions (as in figure 3a). Beal defined
a node as “consistent” if its heuristic
value is the same as the backed-up
value from a 1-ply search over its
is usually less than 0.1 percent and k from the end of the game.)
descendants (Beal 1980, p. 106). Most
seldom reaches 1 percent. Moreover, There is insufficient space to illus-
interestingly, although this definition
even if not a single terminal node is trate the role of tactics more thorough-
was part of a model of quiescence
visited during the search, minimax ly here, but the benefits of minimax
search and forward pruning, Beal did
look-ahead is very useful. look-ahead can also be seen in a more
not relate this model to a model show-
The avoidance of traps in a more general way. The static evaluation
ing pathological behavior presented
general sense is actually one of the functions used in practice usually
earlier in the same article, in the sense
major tasks of minimax look-ahead in incorporate only static knowledge
of using quiescence search and forward
chess: to recognize forced material (that is, there is no attempt to evaluate
pruning to overcome pathology.
gain or loss. Such a search discovers the outcome of immediate impending
Now let us generalize the notion of
“double attacks” (see figure 3b), “piece combat), whereas the dynamic aspects
consistency in order to get a model of
overloading,” “decoying,” and all the are evaluated conveniently by a full-
quiescence that reflects the empirical-
other tactical chess elements about width minimax search. In terms of
ly derived notion.
which the usual static evaluation planning, the static evaluation func-
functions have no explicit knowledge. tions can only measure the current Definition: Given a static evaluation
Because tactics plays a major role in state of goal achievement (for exam- function f returning heuristic values
games like chess, checkers, or kalah, ple, the material advantage of a posi- (HV), a node k is n-ply-quiescent if and
searching deeper is a very successful tion), but cannot realize whether and only if the static heuristic value HV of
method for improving the perfor- how such a goal can be reached (for k assigned by f is equal to the mini-
mance of programs for such games. In example, by forcing the opponent in max value of k resulting from a full-
backgammon, however, forced varia- such a way that material gain cannot width search of the subtree below k up
tions are not very important. Thus be prevented). to depth n, where the terminal nodes
Berliner’s (1980) backgammon pro- It is interesting to note that it is are assigned heuristic values by f.
gram, based primarily on knowledge exactly this difference between static In practice, the heuristic evaluators
and not on search, was very success- and dynamic evaluation, together with usually return a wide range of values.
ful. a fixed search horizon, that causes one Therefore, strict equality would be
The minimax propagation rule of the major defects of the minimax achieved very seldom, and a relax-
backs up exactly one value for each method, the horizon effect; see also ation, for example, in the sense of
position, the one it assumes to be best. Berliner (1973). Berliner (1981, p. 585) “small difference,” would be more
Consequently, this rule is well suited showed that this effect does not realistic. However, to make the model
for games in which each player has influence the played move very often as simple as possible, we will restrict
exactly one move from each position; when searching is sufficiently deep. ourselves here to a two-valued func-
in fact, it is a generalization of the Nevertheless it is worth trying to tion. In this context, the criterion of
game-theoretic relationship. Of course avoid it by investigating important equality seems appropriate.
it is important that this heuristic variations more deeply than others The following model is an attempt
value is sufficiently reliable. Fortu- (Kaindl 1981, 1982, 1983a, 1983b). to concentrate on the strict separation

74 AI MAGAZINE
of the dynamic and static aspects dis- sufficient to capture their dynamic the latter purpose Michon’s criterion
cussed above, using the definition of aspects. seems to fit much better, but there it
n-ply-quiescence. While a relatively Unfortunately, this model has so far models the notion of forcedness rather
complex relationship between the true resisted formal analysis, because of the than quiescence. Although these con-
values and heuristic estimates is intro- relatively complex relationships it cepts are related in some sense, it
duced, the overall model has been contains. Thus, the statement that it seems useful to distinguish between
kept as simple as possible. models observed behavior remains a them, especially for their application
1. The tree structure has a uniform conjecture. Nevertheless, as an illus- in practice. It appears to be safe to use
branching factor b. tration it could help us to approach a the criteria for forcedness also for qui-
2. True values of nodes (TV) are either deeper understanding. escence. For example, the criterion
WIN or LOSS. Michon (1983) investigated the “replies to check” is used successfully
3. True values have the game-theoret- influence of quiescence analysis for- for both concepts in most chess pro-
ic relationship. mally and found it to be beneficial. grams. Responses when the king is in
4. Heuristic values (HV) are either +1 Unfortunately, he considers quies- check are usually small in number,
(estimating WIN) or -1 (estimating cence mainly from a quantitative and consequently the player on move
LOSS). point of view, focusing on the natural
5. Probabilities of error e+ and e- are termination of a quiescence search.
defined as follows (k being a node): However, the criterion for quiescence
e+ (k) = P (HV(k) = +1 | TV(k) = used in his model is not related to the In the future, extensively
LOSS) static evaluation used to model the searching to variable
e- (k) = P (HV(k) = -1 | TV(k) = WIN) empirically derived notion of quies-
6. For all nodes l (those n-ply-quies- cence discussed above. Therefore it is depth probably will
cent according to the definition) and doubtful whether this analysis pro- further improve results
for all nodes m (those not n-ply-quies- vides an explanation for the actually
cent), the following conditions hold: observed benefits of minimaxing. In
e+ (l) < e+ (m) fact, a theorem based on this model
e- (l) < e- (m) indicates that the probability of error
e+ (l) < e- (m) converges toward zero, regardless of is usually forced. Hence, a reliable
e- (l) < e+ (m) how poor the static evaluation is static evaluation of such a position is
7. The number of nodes that are n-ply- (Michon 1983, p. 93). Considering the fairly difficult, and the result is non-
quiescent is small compared to the extreme case in which the static eval- quiescence.
number of nodes that are not n-ply- uation function systematically consid- However, the inverse process is
quiescent. ers all LOSSes as WINs and vice versa, rather dangerous and must be handled
Assumptions 1 to 5 are fairly stan- we can note that such behavior has very carefully. Ken Thompson per-
dard in many models and are given not yet been observed in practice. formed an experiment with the Com-
here for completeness. However, puter World Champion of 1980,
assumptions 6 and 7 are new and BELLE, which can serve as empirical
model the properties that are very Consequences for the evidence here. Thompson changed
likely the ones responsible for the dra- Proper Use of Minimaxing BELLE to count capture moves as only
matic benefits observed in practice. half a ply: Whenever two (four ...) cap-
Assumption 6 introduces the relation- In practice, searching to a strictly fixed ture moves in a path occur, this path
ship between the errors of evaluating depth is rather unusual, as simple cri- is searched one (two ...) ply deeper.
quiescent and nonquiescent nodes. teria for relaxing the depth restric- This procedure does not seem unrea-
The number n of n-ply-quiescence can tion—that is, quiescence (making sonable, considering the enormous
be treated as a parameter of the model, direct capture moves) or forcedness utility of the capture quiescence
and it seems likely that there is a (forced response when the king is in search. However, he changed it back
strong relationship between n-ply-qui- check)—have given very good results. after having had bad experience with it
escence and the benefits of searching In the future, extensively searching to in several games. The reason for this is
the subtree n plies deep. The assump- variable depth probably will further rather simple: Capture moves are very
tion itself, that the heuristic values are improve results, since this effectively useful for refuting bad lines within the
much more reliable for quiescent posi- utilizes the benefits of minimax tree (in the sense of alpha-beta cut-
tions than for nonquiescent ones, is searching while simultaneously reduc- offs), but these refutations became
undoubtedly justified by practical ing its defects. much more expensive as a result of
observations. (Why else are resources Michon (1983, p. 78) used the same this change. The subtrees below the
spent for quiescence searching?) criterion for quiescence and to decide refuting capture moves usually had to
Assumption 7 models the dynamic whether a move should be counted as be searched deeper.
nature of tactical domains such as a ply of depth. (When a move is not Even before these empirical data
chess, checkers, or kalah, where at counted, then the corresponding path were available, a model was created
least up to now, no static evaluation is searched deeper, and in effect, the (Kaindl 1983b) that also includes the
functions have been written that are search horizon becomes variable.) For “interest” of a position in the decision

FALL 1988 75
whether to extend the horizon from it. New York: Association for Computing Computer-Schach. Doctoral diss., Techni-
When comparing estimated values of Machinery. cal Univ. of Vienna.
the interest with the aspiration win- Beal, D. F. 1982. Benefits of Minimax Michon, G. 1983. Recursive Random
dow (represented by the current values Search. In Advances in Computer Chess 3, Games: A Probabilistic Model for Perfect
of and ß), one can try to avoid wast-
a
ed. M. R. B. Clarke, 17-24. Oxford: Perga- Information Games. Ph.D. diss., Univ. of
ing effort on variations that are not mon. California at Los Angeles.
likely to become important for the Beal, D. F. 1980. An Analysis of Minimax. Nau, D. S. 1983a. Pathology on Game Trees
result of the search. In Advances in Computer Chess 2, ed. M. Revisited, and an Alternative to Minimax-
R. B. Clarke, 103-109. Edinburgh, Scotland: ing. Artificial Intelligence 21(1,2): 221-244.
Edinburgh University Press. Nau, D. S. 1983b. Decision Quality as a
Conclusion Berliner, H. J. 1981. An Examination of Function of Search Depth on Game Trees.
Brute Force Intelligence. In Proceedings of Communications of the ACM 30(4): 687-
This discussion is an attempt to help the Seventh International Joint Conference 708.
on Artificial Intelligence, 581-587. Menlo
bridge the gap between theory and Nau, D. S. 1983c. On Game Graph Struc-
Park, Calif.: International Joint Confer-
practice for minimax look-ahead. The ture and Its Influence on Pathology. Inter-
ences on Artificial Intelligence.
arguments in favor of minimaxing and national Journal of Computer and Informa-
Berliner, H. J. 1980. Backgammon Comput- tion Sciences 12(6): 367-383.
the model presented should also sup- er Program Beats World Champion.
port efforts to obtain the desired theo- Artificial Intelligence 14(1): 205-220. Nau, D. S. 1982. An Investigation of the
retical evidence of the benefits Causes of Pathology in Games. Artificial
Berliner, H. J. 1973. Some Necessary Condi- Intelligence 19(3): 257-278.
observed in practice. Unfortunately,
tions for a Master Chess Program. In Pro-
the model resulting from the empirical ceedings of the Third International Joint Nau, D. S. 1980. Pathology on Game Trees:
approach used here has resisted formal Conference on Artificial Intelligence, 77- A Summary of Results. In Proceedings of
analysis up to now because of the rela- 85. Menlo Park, Calif.: International Joint the first National Conference on Artificial
tively complex relationships between Conferences on Artificial Intelligence. Intelligence, 102-104. Menlo Park, Calif.:
the true values and heuristic estimates American Association for Artificial Intelli-
Bratko, I., and Gams, M. 1982. Error Analy-
gence.
of whole subtrees. However, it can sis of the Minimax Principle. In Advances
serve as an illustration of an approach in Computer Chess 3, ed. M. R. B. Clarke, Nau, D. S. 1979. Quality of Decision Versus
Depth of Search on Game Trees. Ph.D.
to deeper understanding. finally, the 1-15. Oxford: Pergamon.
diss., Duke Univ.
insights gained from such investiga- Campbell, M. S., and Marsland, T. A. 1983.
Pearl, J. 1983. On the Nature of Pathology
tions should help to improve practical A Comparison of Minimax Tree Search
in Game Searching. Artificial Intelligence
applications. Algorithms. Artificial Intelligence 20(4):
20(4): 427-453.
347-367.
Pearl, J. 1980. Asymptotic Properties of
Postscript Elo, A. 1978. The Rating of Chess Players,
Minimax Trees and Game-Searching Proce-
Past and Present. London: Batsford.
dures. Artificial Intelligence 14(2): 113-138.
Although this article is devoted to Horacek, H.; Kaindl, H.; and Wagner, M. Schrüfer, G. 1986. Presence and Absence of
minimaxing, we should note that a 1987. Probabilities in Game-Playing: Possi- Pathology on Game Trees. In Advances in
different method of propagating values ble Meanings and Applications. In Proceed- Computer Chess 4, ed. D. F. Beal, 101-112.
has recently become popular in the ings of the Third Austrian Meeting on
Oxford: Pergamon.
context of theoretical analysis of game Artificial Intelligence, 12-23. Berlin:
Springer-Verlag. Shannon, C. E. 1950. Programming a Com-
tree searching. This method considers puter for Playing Chess. Philosophical Mag-
heuristic estimates as independent Kaindl, H. 1983a. Quiescence Search in
azine 41(7): 256-275.
probabilities of winning. It was first Computer Chess. In Computer Game-Play-
ing: Theory and Practice, 39-52. Chich- Slagle, J. R., and Bursky, P. 1968. Experi-
used by Slagle and Bursky (1968) and ments with a Multipurpose, Theorem-Prov-
much later was again proposed by ester, England: Ellis Horwood. Also in
SIGART Newsletter 80, ed. M. A. Bramer, ing Heuristic Program. Communications of
Pearl (1983). There is some evidence the ACM 15(1): 85-99.
that it may be rather useful for games 124-131, 1982.
Kaindl, H. 1983b. Searching to Variable Thompson, K. 1982. Computer Chess
where minimaxing is pathological Strength. In Advances in Computer Chess
Depth in Computer Chess. In Proceedings
(Nau 1983a, 1983c). Unfortunately, no 3, ed. M. R. B. Clarke, 55-56. Oxford: Perga-
of the Eighth International Joint Confer-
suitable tree-pruning procedure (such ence on Artificial Intelligence, 760-762. mon.
as alpha-beta for minimaxing) is avail- Menlo Park, Calif.: International Joint Con- Turing, A. M. 1953. Digital Computers
able. Thus, it appears that for comput- ferences on Artificial Intelligence. Applied to Games. In Faster than Thought,
er chess practice, this method has Kaindl, H. 1982. Dynamic Control of the ed. B. V. Bowden, 286-295. London: Pitman.
more conceptual defects than mini- Quiescence Search in Computer Chess. In
maxing (Horacek, Kaindl, and Wagner Proceedings of the Sixth European Meeting
1987). on Cybernetics and Systems Research, ed.
R. Trappl, 973-978. Amsterdam: North-
References Holland.
Beal, D. F. 1983. Recent Progress in Under- Kaindl, H. 1981. Realisier ung von
standing Minimax Search. In Proceedings langfristigem Planen und Ma[Help - beta
of the ACM Annual Conference, 164-169. sign]nahmen gegen den Horizont-Effekt im

76 AI MAGAZINE

You might also like