Content uploaded by Hamid Reza Nasrinpour
Author content
All content in this area was uploaded by Hamid Reza Nasrinpour on Mar 11, 2015
Content may be subject to copyright.
DEPLOYING FUZZY LOGIC IN A BOXING GAME
Hamid Reza Nasrinpour, Siavash Malektaji, Mahdi Aliyari Shoorehdeli and Mohammad Teshnehlab
Department of Electrical and Computer Engineering
K. N. Toosi University of Technology
Seyedkhandan, Tehran, Iran
nasrinpour, siavashmalektaji@ee.kntu.ac.ir
m_aliyari, teshnehlab@eetd.kntu.ac.ir
KEYWORDS
Fuzzy Logic, Computer Game, Rule-Based System, Boxing
ABST RACT
Nowadays computer games have become a billion dollar
industry. One of the important factors in success of a game is
its similarity to the real world. As a result, many AI
approaches have been exploited to make game characters
more believable and natural. One of these approaches which
has received great attention is Fuzzy Logic. In this paper a
Fuzzy Rule-Based System is employed in a fighting game to
reach higher levels of realism. Furthermore, behavior of two
fighter bots, one based on the proposed Fuzzy logic and the
other one based on a scripted AI, have been compared. It is
observed that the results of the proposed method have less
behavioral repetition than the scripted AI, which boosts
human players’ enjoyment during the game.
INTRODUCTION
The academic definition of Artificial Intelligence (AI) states
that AI is creating machines which can sensibly think and act
as humans (Russell and Norvig 1995). This classic AI is
usually concerned with the optimum solution to a problem
and discusses how this solution can be found. Game AI is a
code or technique which a computer uses to control Non-
Player Characters (NPCs). It does not care how a computer
makes a decision or thinks about the problem, it might use
either a decision tree or a huge database, the point is just its
reaction.
Game balance has been one of the considerable topics in
Game AI. If a computer opponent always just followed
similar patterns, or played too easily or strictly, the game
would be annoying. Furthermore, the aim is not to defeat
players, but rather to keep them amused. In fact, none of the
opponents should be unbeatable.
In Game AI, natural game laws should be followed and
cheating should be prevented as far as possible. Cheating AI
is a term which refers to a situation where the AI has more
information and advantages than the players. This is often
implemented in games to increase the abilities of a machine
and it might be acceptable if the player is not apprised of it.
Some examples of cheating AI can be found in (Scott 2002).
As a matter of fact, applying these cheats represents
weaknesses and limitations of AI against human knowledge.
Therefore it would be better to look for the methods which
represent human knowledge, like Fuzzy rule-based systems,
as they are a widespread form of Fuzzy logic (Zadeh 1965).
Fuzzy logic can easily rate any input based upon importance.
Additionally it is really suitable to do multiple operations at
once (Doss 2006). Human knowledge and the ways in which
humans think and infer can be directly put into a game by
Fuzzy rule-based systems. For instance, in a Fighting Game
(Rollings and Ernest 2006), a game developer can take
advantage of Fuzzy logic to reach higher levels of realism
and human-level AI in behavior of the opponent's character.
This makes the human players feel that they are playing
against another human, not an omniscient computer, which
knows about all of the angles and distances between itself
and its opponent and consequently, makes no mistake unless
it deliberately decides to make the game easier for the human
opponent.
In recent years, we have witnessed many applications of
Fuzzy techniques in the domain of games. The game
S.W.A.T. 2 has employed Fuzzy logic as an instance of action
games (Johnson and Wiles 2001). In the genre of Real Time
Strategy (RTS) games, Civilization: Call to Power uses
Fuzzy State Machines (FuSM) as well (Johnson and Wiles
2001). The great performance of Fuzzy methods in 2009
simulated car racing championship can be found in
(Loiacono et al. 2010). Another Fuzzy controller for a car
racing championship is discussed in (Perez et al. 2009), and
also a Fuzzy-based architecture has been tested in The Open
Car Racing Simulator (TORCS) in (Onieva et al. 2010).
There is a Fuzzy Q-learning method which has been
implemented in the game of Pac-Man (DeLooze and Viner
2009). An agent-based Fuzzy system has been applied in the
Battle City game into the bargain (Li et al. 2004).
In this paper, a method without cheating has been introduced
for applying Fuzzy logic in a fighting game. In order to
avoid cheating and make a natural and human-level AI, first
the game engine was implemented independently. Then the
game engine would give control of the characters to a human
player or an AI in the same way. Moreover, the rules which
the Fuzzy AI fires are the Fuzzy rules which can be followed
easily by a human and do not have any complicated
computational overhead.
BACKGROUND
Boxing is an old Atari 2600 video game released by
Activision group in 1980. Boxing shows a top-down view of
two boxers which can punch their opponent, as shown in
Figure 1. The choice between punching hands (right or left)
is made automatically and the human player just presses the
hit button. Two boxers are always in front of each other and
they can only move up, down, forward and backward. So
they cannot rotate. The game finishes after two minutes or
when one boxer knockouts the other one by giving him 100
punches.
Figure 1: Original Atari Boxing Game
Clever Boxer, discussed in this paper, is an extension of the
original Boxing game where boxers can rotate and go behind
the opponent and the punching hand selection is made
manually. So the human player must press the proper button.
The boxers can make fast movements in normal directions
(forward, backward, left and right) and if they try to punch
while moving fast, their punch on the opponent will be more
powerful. In addition, the boxer must have enough stamina
to punch or to move fast.
GAME ENGINE
The game engine has been designed based on physical rules
in order to be able to be implemented on boxer agent bots.
Hence the game could be considered as an agent-based
simulation software of a real boxer agent's behavior.
Design
This section covers some important rules of the game. The
boxer agent, whose behavior is implemented by a Finite
State Machine (FSM), has three different types of
commandable states which can directly be given by a human
or AI player, which are shown in Table 1. If a boxer is
commanded to punch, it will stand still and throw a punch.
But the commands of rotation and the commands of
direction can be given and run simultaneously. The act of
throwing a punch, implemented by a FSM, takes place in 4
states. In fact, in each state the hand of the boxer reaches out
slightly in a way that in the 4th state the hand is completely
stretched. When a punch hits the opponent, it will be
effective, if and only if the hand is in the 3rd or 4th state.
Besides, the powers of punches in 3rd and 4th states are
different from each other.
There are some other states that are not directly
commandable and a boxer could be in. An expected state of
these non-commandable states is the Overshooting state in
which the boxer is punched in its back or stomach. The
human player or AI does not have any control over the boxer
in Overshooting states. The complete list of non-
commandable states with exact definitions, which implicitly
explain the rules of the game, is shown in Table 2.
Table 1: Commandable States of Boxer Agent
State Type
State Name
Direction
States
1.Fix
2.GoingForward
3.GoingBackward
4.GoingRight
5.GoingLeft
6.GoingForwardFast
7.GoingBackwardFast
8.GoingRightFast
9.GoingLeftFast
Punching
States
1.PunchRight
2.PunchLeft
3.NoPunching
Rotation
States
1.RotateRight
2.RotateLeft
3.NoRotation
Table 2: Non-Commandable States of Boxer Agent
Name
Description
BeingPushed
The boxer is being hustled by the other one. It
happens when a boxer is moving faster than the
other one.
BoxerContact
Two boxers have a collision with each other.
CircularOvershooting
The boxer is punched in its eye, and then it
rotates up to 45 degrees while overshooting.
DoubleOvershooting
If the back of a boxer come into contact with
the ring border while overshooting, it will be in
this state.
FastPunching
The boxer punches while fast moving. This
punch is more powerful than a normal punch.
Overshooting
The boxer is punched in its back or stomach and
the human player or AI cannot control it.
Punching
The boxer punches while normal moving.
RingAccelaration
The boxer pushes its back toward the ring
border to go forward faster. If a boxer punches
in this state, it will go to the FastPunching state.
RingContact
The body of the boxer has touched the ring
border
As it was mentioned earlier, a boxer will be unable to punch
or to move fast if its stamina is lower than a specific level.
Also it will not be able to punch if it receives too many
punches in its hand. In other words, every hand can endure
up to a specific number of hits; otherwise the hand will lose
its ability.
Implementation
The game framework, written in Microsoft Visual C#,
provides a control loop driven by an external timer to handle
animations and collisions. The framework also gives AI an
opportunity to perform its own processing, while human
players can asynchronously command their boxers.
If AI were reliant on the speed and precision of its punches
rather than its playing strategy, it could be seen as cheating
because it is impossible for a human player to replicate the
computer's effort as fast or as easily as the computer.
Therefore the game engine applies two different types of
delay to AI in order to make it use more realistic game
strategies and act like a human; one is the information delay,
which is the delay of receiving data of the environment and
the opponent's character from the game engine. The other
one is the delay in execution after which a command from
the AI is issued. The game engine informs AI of the nature
of these delays.
AI
Two different AIs are implemented to control the boxer
during the game. One is based on scripted AI and the other
one is based on Fuzzy logic. In this section, these two
implementations are discussed.
Scripted AI
Scripting (Bourg and Seeman 2004) is currently the most
common means of control in Game AI. Most developers
resort to scripts to implement Game AI for complex games
where the number of choices at each turn varies from
hundreds to even thousands. Some advantages of scripting
are being understandable, easy to implement and easily
extendable (Tozour 2002).
In this method, some scripts have been written for different
situations which might happen during the game. For
instance, when a boxer is too close to the opponent in a way
that its punches are not effective, it pushes forward the
opponent in a fast movement, or when its hand does not have
enough stamina for punching, it keeps itself at a safe
distance from the opponent, where the opponent’s punches
would not be efficient and looks forward to the timeout.
Furthermore, some parameters have been defined based on
the AI character's and the opponent's health, stamina and
abilities of their hands. A selection of these parameters has
been used in some scripts. A simple example of the
parameters’ effects on the scripts is that the scripted AI will
play its conservative scripts if the AI boxer's health is less
than half of its opponent's health.
Fuzzy Rule-Based AI
Figure 2 provides an overview of the architecture of the
game engine and its relation to the Fuzzy rule-based AI. It is
based on a classical three-layer hierarchy: perception, control
and actuation layer.
Figure 2: Overview of the system architecture
The perception layer, which receives the position, health,
stamina and hand abilities of two boxers from the game
engine, first of all, anticipates the current situation of the
boxers based on the information delay, the execution delay
and its own boxing ring simulator. After that, calculation of
the Fuzzy system inputs like Alpha, Distance and Danger is
performed.
Alpha as shown in Figure 3 represents the degree between
the perpendicular line to the body of the AI boxer, and the
link from the central point of a boxer to the other one.
Distance represents the space between the boxers. Since the
AI boxer tries to escape from the corners of the ring, the
danger of a position should be estimated independently of
the opponent's position. Hence Danger represents how
dangerous the current position of the AI boxer is.
In the control layer, first of all, by using a script, the AI
checks whether its boxer can throw a punch at the opponent
or it should change its position. If its punch does not hit the
opponent, the AI will call its Fuzzy system to navigate the
boxer toward the best position. The main idea of Fuzzy
system is dividing the problem into two sub-problems
including: 1) Finding the opponent and moving toward it. 2)
Adjusting the exact distance and angle for an efficient
punching to the opponent and evading the opponent’s
punches. Herein, a Fuzzy subsystem copes with each sub-
problem and a 3rd Fuzzy subsystem supervises the efficacy of
each of those two subsystems.
As a direct solution to the sub-problems mentioned above,
the Fuzzy system includes three Fuzzy subsystems: 1) Near
Fuzzy System, which specifies the importance of Distance,
the AI boxer's stamina and the distance of the AI boxer's
hand to its opponent and the opponent's hand to the AI
boxer. It is most effective when the AI boxer and its
opponent are close to each other; 2) Far Fuzzy System,
which specifies the importance of Distance, Alpha and
Danger. It is most effective when the AI boxer and its
opponent are far from each other; 3) Overall Fuzzy System
which specifies the importance of the Near Fuzzy System
verdict and the Far Fuzzy System verdict. It should be noted
that all three Fuzzy subsystems operate during the whole
decision making loops and all rules are evaluated in parallel.
Figure 3: Displaying Alpha in a screenshot of the game
In all of the Fuzzy systems, Takagi-Sugeno Fuzzy model is
employed (Takagi and Sugeno 1985). Fuzzy system input
variables are codified by some simple membership functions
as shown in Figure 4 and output membership functions are
shaped with singletons. It is worthy to mention that the
membership functions were selected intuitively at the
beginning. Nevertheless after some preliminary experiments
during the game development, they were tuned for a
smoother game play.
Figure 4: Fuzzy Membership Functions
Near Fuzzy System rules are shown in Table 3, where
columns represent the variable OppStamina and rows
represent the variable MyStamina.
Table 3: Near Fuzzy System Rules
OppStamia
MyStamina
Low
High
Low
W_MyHandDist = 0.3*Attack/5
AND
W_OppHandDist = 0.3*Flee/5
AND
W_Distance = -0.3*Flee/Attack
AND
W_MyStamina = 1
W_MyHandDist = 0.3*5/Flee
AND
W_OppHandDist = 1
AND
W_Distance = -0.5*Flee/5
AND
W_MyStamina = 1
High
W_MyHandDist = 1
AND
W_OppHandDist = 0.3*Attack/5
AND
W_Distance = 0.5*Attack/5
AND
W_MyStamina = 0
W_MyHandDist = Attack/5
AND
W_OppHandDist = Flee/5
AND
W_Distance = 0.3
AND
W_MyStamina = 0.3
As an example in the case where both OppStamina and
MyStamina are low the following rule will be triggered.
IF (MyStamina = Low AND OppStamina = Low) THEN
(W_MyHandDist = 0.3*Attack/5 AND W_OppHandDist
= 0.3*Flee/5 AND W_Distance = -0.3*Flee/Attack AND
W_MyStamina = 1)
Far Fuzzy System rules are shown in Table 4, where columns
represent the variable Distance and rows represent the
variable Alpha.
Table 4: Far Fuzzy System Rules
Distance
Alpha
Short
Long
Low
W_Distance = 0.1
AND
W_Alpha = 0
AND
W_Danger = 1
W_Distance = 1
AND
W_Alpha = 0
AND
W_Danger = 1
Medium
W_Distance = 0 AND W_Alpha = 1 AND W_Danger = 0.7
Large
W_Distance = 0 AND W_Alpha = 1 AND W_Danger = 0
Finally, Overall Fuzzy System rules are as follows:
IF (Distance = Short) THEN (W_NearFuzzySystem =
1 AND W_FarFuzzySystem = 0)
IF (Distance = Long) THEN (W_NearFuzzySystem =
0 AND W_FarFuzzySystem = 1)
IF (Danger = High) THEN (W_NearFuzzySystem = 0
AND W_FarFuzzySystem = 0.8)
IF (Alpha = Large) THEN (W_NearFuzzySystem = 0
AND W_FarFuzzySystem = 0.7)
W_parameter denotes the weight and importance of the
parameter, where the parameter can be one of the variables
MyHandDist, OppHandDist, Distance, MyStamina, Alpha
and Danger; A negative value of W_Distance states to
increase the distance and a positive value states to decrease
it. In addition, W_FarFuzzySystem and W_NearFuzzySystem
are the degrees of truth of the statements, inferred from the
outputs of Far Fuzzy System and Near Fuzzy System
respectively. This is how the Overall Fuzzy System affects
the whole decision making process. Flee and Attack, which
can be configured by user between 1 and 10, specify the
preference of the boxer for escaping or attacking. These two
parameters can be set to 5 for a typical boxer with normal
behavior. Some important notations are summarized in Table
5.
Table 5: Nomenclature
Notation
Short Definition
Alpha
Degree between two boxers
Attack
Tendency of AI boxer to attack
Danger
Danger of being near of the ring border
Distance
Distance between two boxers
Flee
Tendency of AI boxer to escape
MyHandDist
Distance of AI boxer's hand to its opponent
MyStamina
AI boxer's strength to continue fighting
OppHandDist
Distance of the opponent's hand to AI boxer
OppStamina
Opponent's strength to continue fighting
Finally, in the actuation layer, if the previous layer demands
a punch, it will throw a punch based on the abilities of its
hands. Otherwise it should test all possible actions on its
own ring simulator during two clocks. Then a fitness factor
is calculated for each action based on the weighted
parameters, which have been produced by the Fuzzy system.
In fact, the differences in the values of the parameters (such
as ∆Alpha) during simulation are multiplied by their
corresponding W_parameter (like W_Alpha) and divided by
a normalizer coefficient. The sum of these resultant values
determines the fitness of each action. Additionally, in the
actuation layer, there are some tricks to avoid getting stuck
somewhere in the ring, or persisting in just punching in a
row. For instance, if the boxer retains its position near the
ring edges for a while, the AI ignores the current situation of
the game and moves on to increase its distance from the
sides of the ring.
RESULTS
The proposed AI system has been applied to the introduced
game engine. The performances of five Fuzzy systems with
different configurations were assessed against a scripted AI
and 10 simulations have been performed over each one.
Since there is no standard method for gaging the
‘believability’ of game bots (Gorman et al. 2006), only the
match results (wins and losses) are illustrated in Table 6.
Table 6: Game Results of the Fuzzy AI against the Scripted AI
Fuzzy AI Type
Win
(Knockout)
Win
(Timeout)
Lost
(Knockout)
Lost
(Timeout)
Balanced
(Flee=Attack=5)
60%
10%
10%
20%
Defensive
( Flee = 10,
Attack = 1 )
0%
10%
60%
30%
Offensive
( Flee = 1,
Attack = 10 )
80%
0%
20%
0%
Fairly Defensive
( Flee = 7,
Attack = 3 )
60%
10%
20%
10%
Fairly Offensive
( Flee = 3,
Attack = 7 )
70%
20%
10%
0%
Based on the empirical perception about the AI agents, the
most significant difference between the playing methods of
the Fuzzy AI and the scripted AI is the diverse behavior of
the Fuzzy AI. This matter becomes more obvious when two
scripted AIs play against each other. In this case, it is seen
that many repetitive situations happen in each run of the
game. But when one player utilizes the Fuzzy AI, the game
gets more unpredictable; In fact, the Fuzzy agent insists less
on a specific action and has better adaptation to the game.
CONCLUSIONS
By using a Fuzzy system, rules and membership functions
designed by a human expert, we can avoid cheating and just
follow the natural laws of the game. This could be very
useful for extending and making better games, i.e. when a
human finds a new rule which can be useful in a game, we
will be able to easily insert the new rule into the game as a
Fuzzy rule.
The main benefit of the Fuzzy logic approach to game
development is human-like behaviors, which guarantee
believable and natural behavior (not necessarily perfect) and
increase the satisfaction of human players. Another
advantage of using Fuzzy logic is that, unlike scripted AI, a
developer does not need to consider all possible situations in
a game. Furthermore, the idea of various Fuzzy subsystems
divides the state space of the problem, which makes the
scaling to more demanding settings easy and could be
efficiently exploited in other genres of games. As an
example, in RTS games as the most AI challenging type of
games, two simultaneous main concerns are consuming the
resources for building different structures and training the
army forces for attacking the opponent. To fulfill this
purpose, two Fuzzy systems can be used for handling these
two problems. In addition, a third Fuzzy system could be
used for determining the priority of each of those Fuzzy
systems, according to the current situation of the game.
Future enhancements are required to overcome the deliberate
delay, which is applied to the AI; a simple solution would be
using the probabilistic Bayesian methods in order to predict
the current situation of the opponent. Another possible
outcome of the Clever Boxer is a boxing trainer, just like a
human boxing coach; The AI system watches a boxing game
and then suggests how the player could improve its
performance based on its own Fuzzy human-like rules.
REFE RENC ES
Bourg D. M. and Seeman G. 2004. AI for Game Developers.
O'Reilly, chap. 8.
DeLooze L. L. and Viner W. R. 2009. “Fuzzy Q-Learning in a
Nondeterministic Environment: Developing an Intelligent Ms.
Pac-Man Agent”. In IEEE Symposium on Computational
Intelligence and Games (CIG’09), pp. 162-169.
Doss P. 2006. “Artificial Intelligence in Computer Games”. Bowie
State University, M. Sci. Thesis.
Johnson D. and Wiles. J. 2001. “Computer Games with
Intelligence”. In Proceedings of the 10th IEEE Conference on
Fuzzy Systems (Dec. 2-5), vol. 3, pp. 1355- 1358.
Gorman B. et al. 2006. “Believability Testing and Bayesian
Imitation in Interactive Computer Games”. In Proceedings of
the 9th Int. Conf. on the Simulation of Adaptive Behavior
(SAB'06), Springer, vol. LNAI.
Li Y. et al. 2004. “Fuzzy Logic in Agent-Based Game Design”. In
Proceedings of the Annual Meeting of the North American
Fuzzy Information Processing Society (NAFIPS '04), pp. 734-
739.
Loiacono D. et al. 2010. “The 2009 Simulated Car Racing
Championship”. In IEEE Transactions on Computational
Intelligence and AI in Games, vol. 2, no. 2, pp. 131-147.
Onieva E. et al. 2010. “Overtaking Opponents with Blocking
Strategies Using Fuzzy Logic”. In IEEE Conference on
Computational Intelligence and Games (CIG’10), pp123-130.
Perez D. et al. 2009. “Evolving a Fuzzy Controller for a Car Racing
Competition”. In IEEE Symposium on Computational
Intelligence and Games (CIG’09), pp. 263-270.
Rollings A. and Ernest A. 2006. Fundamentals of Game Design.
Prentice Hall, chap. 13.
Russell S. and Norvig P. 1995. Artificial Intelligence: A Modern
Approach. Prentice Hall.
Takagi T. and Sugeno M. 1985. “Fuzzy identification of systems
and its application to modeling and control”. In IEEE
transaction on systems, man and cybernetics, vol. 15, no. 1, pp.
116-132.
Tozour P. 2002. The Perils of AI Scripting
.
In Rabin S. AI Game
Programming Wisdom, Charles River Media, pp. 541-547.
Zadeh L. 1965. “Fuzzy sets”. Inf. Control, vol. 8, pp. 338-353.
