OPPONENT MODELING IN REAL

OPPONENT MODELING IN REAL
OPPONENT MODELING IN REAL-TIME STRATEGY GAMES
Frederik Schadd, Sander Bakkes and Pieter Spronck
Universiteit Maastricht
MICC-IKAT
P.O. Box 616
NL-6200 MD Maastricht
The Netherlands
e-mail: [email protected], {s.bakkes,p.spronck}@micc.unimaas.nl
ABSTRACT
Real-time strategy games present an environment in
which game AI is expected to behave realistically. One
feature of realistic behaviour in game AI is the ability
to recognise the strategy of the opponent player. This is
known as opponent modeling. In this paper, we propose
an approach of opponent modeling based on hierarchically structured models. The top-level of the hierarchy
can classify the general play style of the opponent. The
bottom-level of the hierarchy can classify specific strategies that further define the opponent’s behaviour. Experiments that test the approach are performed in the
RTS game Spring. From our results we may conclude
that the approach can be successfully used to classify
the strategy of an opponent in the Spring game.
INTRODUCTION
In computer gaming, real-time strategy (RTS) is a genre
of simulated wargames which take place in real time. In
RTS games, the player needs to construct a base and
build units for the purpose of destroying the opponent.
The opponent is either a human player, or a player controlled by an artificial intelligence (AI). Each unit-type
has particular strengths and weaknesses. To effectively
play an RTS game, the player has to utilise the right
units in the right circumstances.
An important factor that influences the choice of strategy, is the strategy of the opponent. For instance, if one
knows what types of units the opponent has, then typically one would choose to build units that are strong
against those of the opponent. To make predictions
about the opponent’s strategy, an AI player can establish an opponent model. Many researchers point out
the importance of modelling the opponent’s strategy
[2, 3, 9, 10, 12, 14], and state that opponent models
are sorely needed to deal with the complexities of stateof-the-art video games [8].
Establishing effective opponent models in RTS games,
however, is a particular challenge because of the lack of
perfect information of the game environment. In classical board games the entire board is visible to the player;
a player can observe all the actions of the opponent.
Hence, assessing the opponent’s strategy and building
an opponent model is possible in principle, for instance
by using case-based reasoning techniques [1]. In RTS
games, however, the player has to deal with imperfect
information [5]. Typically, the player can only observe
the game map within a certain visibility range of its
own units. This renders constructing opponent models in an RTS game a difficult task. In this paper we
will investigate to what extent models of the opponent’s
strategy can be established in an imperfect-information
RTS-game environment.
The outline of this paper is as follows. We will first introduce the concept of opponent modeling. Then, our
approach to establish effective opponent models in RTS
games will be discussed. Subsequently, our implementation of the approach will be presented. The experiments
that test our approach are described next, followed by a
discussion of the experimental results. Finally, we provide conclusions and describe future work.
OPPONENT MODELING
In general, an opponent model is an abstracted description of a player or a player’s behaviour in a game [8].
Opponent modeling can be seen as a classification problem, where data that is collected during the game is
classified as one of the available opponent models. A
limiting condition is the fact that in RTS games, these
classifications have to be performed in real-time, while
many other computations, such as rendering the game
graphics, have to be performed in parallel. This limits the amount of available computing resources, which
is why only computationally-inexpensive techniques are
suitable for opponent modeling in RTS games.
Preference-based modeling is a commonly used
computationally-inexpensive technique [4]. The technique identifies the model of an opponent by analyzing
the opponent’s choices in important game states. Due
to the visibility limitations in RTS games, however, it is
common that choices of the opponent cannot always be
observed.
In the present research we use Spring, illustrated in
IMPLEMENTATION
Figure 1: Screenshot of the Spring game environment.
In the screenshot, airplane units are flying over the terrain.
Figure 1, which is a typical and open-source RTS game.
A Spring game is won by the player who first destroys
the opponent’s ‘Commander’ unit. We used the freelyavailable ‘AAI’ artificial intelligence player [15] to combat opponents.
APPROACH
A straightforward approach to opponent modeling is the
following. First, a number of possible opponent models is established, then the confidence level of each opponent model is calculated, and finally, the opponent
model with the highest confidence level is selected. An
enhancement of this approach is to apply an hierarchical
ordering on the possible opponent models [7]. This hierarchical approach allows the division of a relatively complex classification problem, into several relatively simple
classification problems. In addition, the hierarchical approach makes it possible to use different classification
methods in each level or the hierarchy. For establishing
opponent modeling in RTS games, we follow the hierarchical approach.
This section discusses our implementation of hierarchical opponent modeling in RTS games. In our implementation, we establish a hierarchy consisting of two levels.
The top-level of the hierarchy classifies the opponent’s
general play style. The bottom-level of the hierarchy
classifies the opponent’s choice of units built.
In the Spring game we discriminate between an aggressive, and a defensive play style. For an aggressive play
style we discriminate at the bottom level between predominantly using the following four unit-types: (1) KBots, (2) Tanks, (3) Ships, and (4) Airplanes. Each unittype has specific strengths and weaknesses, and is therefore used to execute a particular strategy. For instance,
K-Bots are relatively fragile but can cross mountains,
and are therefore useful for a strategy against an opponent which attempts to exploit chokepoints between
mountains. Tanks can only manoeuvre on plain terrain
but are relatively sturdy, and are therefore useful for
a strategy against an opponent who constructs strong
defenses.
For a defensive play style we discriminate at the bottom
level between the following three building preferences:
(1) Super Weapon, (2) Tech, and (3) Bunker. These
three building preferences are commonly observed in actual Spring games.
Figure 2 displays the hierarchy of the opponent models.
The hierarchy defines the following strategies.
• Aggressive→K-Bot. The opponent will attack early
and will typically use K-Bot robots.
• Aggressive→Tanks. The opponent will attack early
and will typically use tanks.
• Aggressive→Air. The opponent will attack early
and will typically use airplanes.
Our opponent models will describe the strategy of a
player. We define a strategy as the general play style
combined with the player’s choice of units built. The
most defining element of an opponent’s strategy is the
general play style. We therefore place the general play
style at the top of the hierarchy. Each play style has its
own subcategories that further define behavioural characteristics.
For instance, if it is known that the opponent follows an
aggressive general play style, a logical response would be
to improve one’s defenses. If also the opponent’s choice
of units is known, the defenses can be specialised to be
effective against those specific units.
Figure 2: Hierarchy of the opponent models.
• Aggressive→Sea. The opponent will attack early
and will typically use ships.
• Defensive→Super weapon. The opponent will attempt to construct a super weapon (e.g., a ballistic
missile).
• Defensive→Tech. The opponent will attempt to
reach a high technology level in order to have quick
access to superior units.
• Defensive→Bunker. The opponent will construct a
massive wall of static defenses around his base so
that he has time to construct an army.
Top-level Classifier
A way of performing opponent-model classifications in
a computationally inexpensive fashion, is by using fuzzy
models [16]. Fuzzy models create models of several classifiable classes based on a single numerical feature. The
choice of the numerical feature is crucial, as it should
allow to discriminate between the defined classes.
In our hierarchy, the top level classifier has to discriminate between the classes ‘aggressive’ and ‘defensive’.
An aggressive player typically will spend a large part
of game time attacking. A defensive player, on the
other hand, typically will use most of the game time
for preparing an army, and only needs a small amount
of the game time for an actual attack. As a result, an
appropriate numerical feature to discriminate between
the two classes would be the relative amount of game
time that the opponent spends attacking.
We define an attack as the observed loss of a player’s
own units. When a loss of units is detected, then the
time span around this moment can be regarded as the
time that an attack took place. Because information
about a player’s own units is always available, this definition is suitable for use in an imperfect information
environment.
An example of a top-level player model is shown in Figure 3. The figure illustrates the confidence of the opponent following an aggressive and defensive play style,
as a function of the percentage of game time spent on
attacking.
Theoretical Background
The concept of discounted rewards origins from the principle of repeated matrix-games [6]. A player can make a
choice between several actions and depending on the effectiveness of an action, a reward is received. In the field
of repeated matrix-games, the most important considerations are how to select the initial strategy and how
to evaluate if a deviation from the initial strategy is
needed. A strategy in repeated matrix-games can be a
simple sequence of actions, which is played repeatedly.
It can also include certain rules for cases that take the
actions of the enemy into consideration. In order to determine whether a deviation from the original strategy
is feasible, the expected rewards of the game with and
without deviation are calculated. Here the discount factor is applied, since it is assumed that a player prefers
to receive a reward early in time rather than in the distant future. Hence, rewards in the future are valued
less. This valuation is expressed by multiplying the future reward with the discount factor. The more distant
the reward is, the more often it is multiplied with the
discount factor. If δ is the discount factor and πt is the
reward received at time t, then the expected reward can
be computed as follows [6]:
(1 − δ) ∗
∞
X
πi ∗ δ i−1
When the expected rewards are calculated, a selection
mechanism typically will select the strategy with the
highest expected reward.
Applied Modification
Analogously to calculating the expected rewards, for discriminating between strategies we need to calculate the
confidence value of the opponent applying a particular
strategy. The confidence value is calculated on the basis
of observations during game events. In classical matrixgames, a game event would be one move of both players.
Bottom-level Classifier
The bottom-level classifier has to discriminate between
the subcategories that further define behavioural characteristics. The dynamic nature of RTS games implies
that a player’s strategy may change during the game.
Therefore, the bottom-level classifier needs to emphasize recent event more than past events. To achieve
this, the principle of discounted rewards is applied.
(1)
i=1
Figure 3: Example of a top-level player model.
Since RTS games do not operate in moves, the term
of an event must be redefined. When playing against
an aggressive opponent, the occurrence of an attack is
a suitable event since the player can then observe the
army of the opponent.
When playing against a defensive opponent, however,
an attack is not a suitable event because a defensive
opponent will rarely attack and hence not many observations can be made. Also, waiting for an attack of
a defensive opponent is typically an unsuccessful game
strategy. When playing against a defensive opponent,
it is typical that the player will ‘scout’ around the base
of the opponent. Since scouting will provide improved
information about the opponent’s actions, a scout event
is a suitable moment for calculating confidence values
against a defensive opponent.
When an event is detected, the confidence values of each
possible strategy will be updated according to the observed game information. If δ is the discount factor, ψs,t
the belief that the opponent uses strategy s at event t,
ranging between 0 and 1, π the total reward added at
each event and i the most recent event, then the confidence cs that the opponent uses strategy s is computed
as follows:
cs =
0
X
ψs,t ∗ π ∗ δ i−t
(2)
t=i
The parameter ψs,t is acquired by inspecting all visible units and structures during event t. Each unit
or structure has a value representing a tendency to a
certain strategy. The unit-tendency values were determined by the experimenter, using his own knowledge of
the game [13]. To give three examples of unit-tendency
values: (1) a common defensive tower has a relatively
small tendency towards an opponent using the defensive bunkering strategy, (2) a super-weapon building
has a relatively high tendency towards the defensivesuper-weapon strategy, and (3) an amphibious tank has
a tendency towards both the aggressive tank and the
aggressive sea strategy.
EXPERIMENTS
This section discusses the experiments that test our approach. First we test the top-level classifier, and then
the bottom-level classifier. We finish the section by providing a summary of the experimental results.
Top-level Experiment
As discussed in the previous section, the top-level classifier requires the AI to detect if an attack took place.
This is implemented as follows. The AI will register
all visible units every N seconds, where N is the size
of the time window. The detection algorithm will scan
each frame for lost units. If the amount of lost units
Threshold
10%
15%
20 seconds
94,92166
99,99083
Time Window
30 seconds 40 seconds
99,2271
93,64321
99,57988
96,22472
Table 1: Average defensive confidence values against a
defensive opponent
Threshold
10%
15%
20 seconds
45.10261
33.18209
Time Window
30 seconds 40 seconds
76.27533
57.83488
52.74819
53.14834
Table 2: Average aggressive confidence values against
an aggressive opponent
is above a certain threshold, then each frame inside the
analysed time window is labeled as a moment in which
the opponent was attacking.
Two parameters determine the accuracy of the attackdetection algorithm. The first parameter is the size of
the time window. The second parameter is the unit
threshold, which is a percentage value. We will first
analyse the sensitivity of the parameter settings on the
obtained confidence values. Next, we will discuss the
obtained confidence values as a function over the game
time.
Sensitivity Analysis
For each configuration of parameters, the AI was
matched ten times against an aggressively playing AI
and ten times against a defensively playing AI. As opponent the ‘NTAI’-AI [11] was chosen, since it is relatively straightforward to implement one’s own strategy
into this AI. For the matches where an aggressive opponent is required, the ‘NTAI’-AI was configured with an
aggressive strategy. Analogously, for the matches where
a defensive opponent is required, the AI was configured
with a defensive strategy. For the time window, the
sizes of 20, 30 and 40 seconds were tested. For the unit
threshold the percentages with value 10% and 15% were
tested.
For each match played, the aggressive and defensive confidence values were recorded at each time point. When
the match was played, the average confidence values of
the match were calculated. Note that as in the first five
minutes of a game a player rarely performs an action
that would reveal his strategy, these first minutes were
not taken into account.
Table 1 shows the defensive confidence values of matches
against an defensive opponent, and Table 2 shows the
aggressive confidence values of matches against an aggressive opponent.
The obtained confidence values reveal that all configurations of the top-level classifier perform well when
Figure 4: Average confidence value over time against an
aggressive opponent
Figure 5: Average confidence value over time against a
defensive opponent
Bottom-level Experiment
recognizing defensive players. The best configuration
obtained a confidence value of nearly 100%. When recognizing aggressive players, the best configuration, with
as parameters a units-lost threshold of 10% and a time
window of 30 seconds, obtained a confidence value of
76%. These obtained configurations will be used for the
remainder of this research.
Confidence Value over Time
Using the obtained parameter values, we match the AI
fifty times against an aggressive opponent, and fifty
times against a defensive opponent. We again used the
‘NTAI’-AI as an opponent. For each game, we record the
confidence values as a function over time. The average
confidence values of all test games against an aggressive
opponent are displayed in Figure 4.
In the figure we observe that the average confidence
value is low in the beginning of the game. This is due
to the fact that the opponent is hardly able to attack
at this stage of the game, since he needs to construct
a base first. Therefore, one can safely disregard the
confidence values of the beginning of the game. After
approximately seven minutes of game time, the average
confidence value increases until it stabilizes at approximately 85%.
A similar effect can be observed when examining the
average confidence value over time of the games against
a defensive opponent, as displayed in Figure 5. In the
beginning of the game, the confidence values are nearly
100%. This is because the enemy does not attack in
the beginning of the game. The top-level classifier will
therefore respond with the maximum defensive confidence value during this game stage. One observes that
after about six minutes of game time, the average confidence value stabilizes between 96% and 97%.
For testing the bottom-level classifier, the ‘NTAI’-AI has
been configured such that it resembles each of the specific aggressive opponent models. However, since NTAI
is not able to adequately play the defensive strategies, a
human opponent was chosen to play against the ‘AAI’AI, by following the defensive strategies. For each opponent model, ten experimental matches have been performed. A correct classification of the top-level classifier
is assumed for this experiment. For the discounted reward algorithm, the parameters were set to δ = 20%
and π = 0.8 by the experimenter. We test the bottomlevel classifier’s ability to discriminate between the KBot and tanks aggressive sub-model, and between the
bunker, tech and super weapon defensive sub-model.
Aggressive Opponent
Figure 6 shows the average K-Bot confidence over time
of an opponent using the aggressive K-Bot strategy as
well as the average Tank confidence over time of an opponent using the aggressive tank strategy. It is observed
that both confidence values eventually approximate a
value over 90%. We note that the average confidence of
the aggressive tank-strategy increases more slowly and
at a later stage than the average confidence of the aggressive K-Bot-strategy. This can be explained by the
fact that producing tanks requires more resources, and
therefore more game time will be needed to attack with
tanks.
Defensive Opponent - Bunker
Figure 7 displays the confidence values obtained against
an opponent using the defensive bunker-strategy. We
observe that the bunker confidence rating increases
rapidly after approximately five minutes of game-time.
Over time the obtained average confidence value is 83%.
The instabilities that occur after 35 minutes of gametime can be explained by the fact that at this moment
the AI has discovered structures that may also be used
Figure 6: Average confidence value over time for the
aggressive sub-models.
by an opponent using the tech-strategy.
Figure 8: Average confidence value over time for an opponent using the defensive tech-strategy.
Defensive Opponent - Super Weapon
Figure 9 displays the confidence values obtained against
an opponent using the defensive super-weapon-strategy.
Analogously to the results obtained against an opponent using the defensive tech-strategy, we observe that
the confidence values of the bunker-strategy are higher
than the confidence values of the super-weapon-strategy.
After approximately 25 minutes the confidence value of
super-weapon-strategy steadily increases. After approximately 41 minutes, the discounted-reward algorithm
temporarily decreased the confidence value because the
AI did not observe super-weapon structures any more.
Eventually the bottom-level classifier obtains a confidence value of 75%.
Figure 7: Average confidence value over time for an opponent using the defensive bunker-strategy.
Defensive Opponent - Tech
Figure 8 displays the confidence values obtained against
an opponent using the defensive tech-strategy. We observe that for the largest part of the game, the confidence values of the bunker-strategy are higher than the
confidence values of the tech-strategy. This can be explained as follows. First, we observed that scout units
were destroyed before they were able to scout the base
of the opponent. Second, high level units and structures
that define the tech-strategy can only be constructed in
later stages of the game. This implies that in the earlier
stages only structures that belong to a different strategy can be observed. When at a later stage of the game
the AI is able to observe structures that are characteristic for the tech-strategy, the confidence value of the
tech-strategy increases.
Figure 9: Average confidence value over time for an opponent using the defensive super-weapon-strategy.
Summary of the Experimental Results
Experimental results obtained with the top-level classifier show that the top-level classifier can accurately discriminate between an aggressive and a defensive player.
Experimental results obtained with the bottom-level
classifier show that the bottom-level classifier can accurately discriminate between the established sub-models
in later stages of the game. In early stages of the game,
the bottom-level classifier was not always able to accurately discriminate between the established sub-models.
This is discussed next.
DISCUSSION
Opponent modeling will typically be implemented in an
actual RTS game for the purpose of automatically classifying the strategy of the opponent. Ideally, an accurate
classification of the opponent’s strategy is available relatively early in the game, at a time when the player is
still able to counter the opponent’s strategy.
In the experiments that test our approach, we observed
that the bottom-level classifier was not always able to
accurately discriminate between the established submodels in an early stage of the game. This phenomenon
can be explained by the fact that, typically, the AI cannot directly observe units and structures that are characteristic for a particular bottom-level strategy. To observe these units and structures, the AI relies on scouting.
A straightforward approach to achieve improved results,
therefore, is to adapt the AI’s scouting behaviour dependent on the need for information of the opponent’s
activities. For instance, in competition against an aggressive opponent, scouting is relatively unimportant.
In competition against a defensive opponent, however,
intensive scouting is vital. Analogously, to emphasise
the information obtained during a scout event, one may
choose to adapt the parameters of the delayed reward
algorithm dependent on the top-level classification.
CONCLUSION
In this paper we proposed an approach for opponent
modeling in RTS games. In the approach, a hierarchical opponent model of the opponent’s strategy is established. The top-level of the hierarchy can classify
the general play style of the opponent. The bottomlevel of the hierarchy can classify strategies that further
define behavioural characteristics of the opponent. Experiments to test the approach were performing in the
RTS game Spring. Our experimental results show that
the general play style can accurately be classified by the
top-level of the hierarchy. Additionally, experimental
results obtained with the bottom-level of the hierarchy
show that in early stages of the game it is difficult to
obtain accurate classifications. In later stages of the
game, however, the bottom-level of the hierarchy will
accurately classify between specific strategies of the opponent. From these results, we may conclude that the
approach for opponent modeling in RTS can be successfully used to classify the strategy of the opponent while
the game is still in progress.
For future work, we will incorporate a mechanism to
adapt the scouting behaviour of the AI dependent on the
top-level classification of the general play style of the opponent. Subsequently, we will investigate to what extent
the classification of the opponent’s strategy can be used
to improve the performance of a computer-controlled
player, and will investigate how new models to describe
the opponent’s strategy can be established automatically.
ACKNOWLEDGEMENTS
The authors wish to extend their gratitude to Marcel
Ludwig and Alexander Miesen for being the human opponent in the experimental games.
This research was funded by a grant from the Netherlands Organization for Scientific Research (NWO grant
No 612.066.406).
REFERENCES
[1] Agnar Aamodt and Enric Plaza. Case-based reasoning : Foundational issues, methodological variations, and system approaches. AI Communications,
7(1), March 1994.
[2] Bruce Abramson. Expected outcome: A general
model of static evaluation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 12:182–
193, 1990.
[3] Hans Berliner. Search and knowledge. In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI 77), pages 975–979, 1977.
[4] Jeroen Donkers and Pieter Spronck. Preferencebased player modeling. In S. Rabin, editor, AI Programming Wisdom 3, chapter 8.4, pages 647–659.
Charles River Media, 25 Thomson Place, Boston,
Massachusetts 02210, 2006.
[5] Robert Gibbons. A Primer in Game Theory, chapter 2.3B. Pearson Education Limited, Edingburgh
Gate, Harlow, Essex CM20 2JE, England, 1992.
[6] Robert Gibbons. A Primer in Game Theory, chapter 2. Pearson Education Limited, Edingburgh
Gate, Harlow, Essex CM20 2JE, England, 1992.
[7] Ryan Houlete. Player modeling for adaptive games.
In S. Rabin, editor, AI Programming Wisdom 2,
chapter 10.1, pages 557–566. Charles River Media, 10 Downer Avenue, Hingham, Massachusetts
02043, 2004.
[8] Jaap van den Herik, Jeroen Donkers, and Pieter
Spronck. Opponent modelling and commercial
games. In Proceedings of the IEEE 2005 Symposium on Computational Intelligence and Games
(eds. Graham Kendall and Simon Lucas), pages 15–
25, 2005.
[9] Donald Knuth and Ronald Moore. An analysis of alpha-beta pruning. Artificial Intelligence,
6(4):293–326, 1975.
[10] Richard Korf. Generalized game trees. In Proceedings of the International Joint Conference on
Artificial Intelligence (IJCAI 89), pages 328–333.
Detroit, MI, August 1989.
[11] Tom Nowell. AI:NTAI. Creator of the game AI
‘NTAI’, http://spring.clan-sy.com/wiki/AI:NTAI,
2007.
[12] Arthur Samuel. Some studies in machine learning
using the game of chechers, ii - recent progres. IBM
Journal, 11:601–617, 1967.
[13] Frederik Schadd. Hierarchical opponent models for
real-time strategy games. Bachelor thesis. Universiteit Maastricht, The Netherlands, 2007.
[14] Jonathan Schaeffer, Joseph Culberson, Norman
Treloar, Brent Knight, Paul Lu, and Duane
Szafron. A world championship caliber checkers
program. Artificial Intelligence, 53:273–289, 1992.
[15] Alexander Seizinger. AI:AAI. Creator of the game
AI ‘AAI’, http://spring.clan-sy.com/wiki/AI:AAI,
2006.
[16] Michael Zarozinski. An open-fuzzy logic library. In
S. Rabin, editor, AI Programming Wisdom, chapter
2.8, pages 90–101. Charles River Media, 20 Downer
Avenue, Suite 3, Hingham, Massachusetts 02043,
2002.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement