Level of Detail based AI Adaptation for Agents in Video Games

Ghulam Mahdi, Yannick Francillette, Abdelkader Gouaich, Fabien Michel and Nadia Hocine

LIRMM, Universit

e Montpellier II, UMR 5506 - CC 477, 161 Rue Ada, 34095 Montpellier Cedex 5, France

Keywords:

Agents, Game AI, Level of Detail, Adaptation, MAS, Organization, Video games, Frame Rate.

Abstract:

This paper suggests multi-agent systems (MASs) for implementing game artiﬁcial intelligence (AI) for video

games. One of main hindrances against using MASs technology in video games has been the real-time con-

straints for frame rendering. In order to deal with the real-time constraints, we introduce an adaptation-oriented

approach for maintaining frame rate in acceptable ranges. The adaptation approach is inspired from the level

of detail (LoD) technique in 3D graphics. We introduce agent organizations for deﬁning different roles of

agents in game AI. The computational requirements of agent roles have been prioritized according to their

functional roles in a game. In this way, adapting computational requirements of game AI works as a means

for maintaining frame rate in acceptable ranges. The proposed approach has been evaluated through a pilot

experiment by using a proof of concept game. The pilot experiment shows that LoD based adaptation allows

maintaining frame rate in acceptable ranges and therefore enhancing the quality of service.

1 INTRODUCTION

Video games are considered nowadays as a main-

stream entertainment and cultural industry. Once a

small and focused discipline, it has turned into a large

industry with applications in education and training.

As a simpliﬁcation, a video game can be consid-

ered as a computational simulation of a set of game

characters that interacts with one or several players.

Each individual game character achieves particular

task in a game and the overall game environment is

governed by rules of the game.

Until recently graphical rendering has been con-

sidered as the Holy Grail for this sector. The main

purpose has been to increase quality and natural-

ness of graphics. Research on game artiﬁcial intel-

ligence ( or behavior of game characters) was not

on the agenda since simple models based on ﬁnite

state machines were considered sufﬁcient to meet re-

quirements of gameplay and player experience. The

development of successful titles such as Neverwin-

ter Nights (BioWare, 2002), Black and White (Stu-

dios, 2001), Max Payne (Remendy, 2001) and Left

4 Dead 2 (Valve, 2008) have challenged this suppo-

sition. These titles, among others, have introduced

more complex, organized and adaptive behaviors by

using more sophisticated artiﬁcial intelligence tech-

niques. In other words, they have clearly demon-

strated the beneﬁts of moving beyond simple behav-

iors.

It is worth noting, that despite using the term “arti-

ﬁcial intelligence” in game sector and academia, this

term refers to different concepts. In fact, the term AI

in games often refers to any model and algorithm that

steer behaviors of game entities. On the other hand,

Artiﬁcial Intelligence (AI) as a scientiﬁc discipline

studies and designs intelligent agents, where an agent

is deﬁned as a system that perceives its environment

and takes actions that maximize its chances of success

(Russell and Norvig, 2010). Despite having different

constraints and objectives, collaboration between the

game sector and AI community has been illustrated

in several works such as (Rabin, 2002; Charles, 2007;

Millington and Funge, 2009). Despite these improve-

ments, AI in the games still remained largely sim-

ple and basic one. As a result, usually ﬁnite state

machines (FSMs) and other simpler algorithms have

been used for modeling behaviors of game characters

(Niederberger, 2005).

According to Neiderberger et al. (Niederberger

and Gross, 2005) current game characters exhibit be-

haviors that are too rudimentary to seem realistic

ones. This may decrease the player sense of alief

(Clark, 2008) that is an important feature to enable

players’ immersion in the game world. Orkin (Orkin,

2006) also states that current game characters lack

ﬂexibility and there is need of more ﬂexible planning

for complex behaviors in game characters. In addition

to that, depending on the nature and type of a game,

there can be hundreds of secondary game characters

182

Mahdi G., Francillette Y., Abdelkader G., Michel F. and Hocine N..

Level of Detail based AI Adaptation for Agents in Video Games.

DOI: 10.5220/0004261501820194

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 182-194

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

that may face competition over getting computational

resources for executing their actions.

Multi-agent systems (MAS) have been growing

prevalence and increased usage in different research

and applied domains. One can notice that there is a

natural mapping between core concepts of MAS and

games. In fact, agents are micro units in MAS that ex-

ecute own behavior to reach their goals. The overall

function of the system is observed at the macro level

as the interaction among all entities of the micro level.

Thus, conceptually each game character can be con-

sidered as an agent and the overall game as a multi-

agent system interacting with one or several players.

Dignum et al. (Dignum et al., 2009b) argue for

using the potential of MASs by incorporating agents

into games for improving game AI. In order to get

an idea of how closely MAS and games are related,

one can observe that apart from conference papers,

since 2009 there has been an annual workshop in ma-

jor academic conferences such as AAMAS (Dignum

et al., 2009a; Dignum, 2011; Dechesne et al., 2012)

and ECAI 2012.

On the other hand, (Dignum et al., 2009b) note

that modeling of game characters around agents may

require relatively more resources so it may not meet

real-time constraints of game engines where funda-

mental concern is rendering of frames with an accept-

able frequency. This is one of the major challenges

that prevents using agents in games.

Here in this paper, we address the question of how

to build game character with agents without deteri-

orating player experience metrics and constraints of

game quality of service (QoS). We propose modeling

of game characters around agents and their dynamical

adaptation to the computational resources for main-

taining acceptable frame rate. The rationale behind

our approach is to use the organization of MAS as a

means to express different priorities to agents depend-

ing on their role. Whenever the quality of game is

not met, agents with less important roles are asked to

adopt simpler behaviors. Consequently, agents with

central roles can continue executing their behavior

and the overall player experience is not affected.

We experimentally show that our approach main-

tains an acceptable player experience when compared

with a simple MAS approach without behavior adap-

tation.

The rest of the paper is organized as follows. Sec-

tion 2 discusses the motivations for the work. Section

3 is about introducing Level of Detail (LoD) tech-

nique, its foundations in computer graphics and its

recent usage in game AI. In section 4, we discuss

our proposed approach, where we present LoD as an

adaptation technique for game AI. Section 4 describes

the experimental framework and the game which is

used to evaluate performance of our approach; then in

section 5, we discuss the results of our experiments

and prototype results of user evaluation. Section 6

discusses some relevant works in the domain and ﬁ-

nally, we conclude the paper in Section 7.

2 MOTIVATIONS

2.1 Behavior Modeling around Agents

Schreiner (Schreiner, 2003) points out the lack of pro-

activeness in case of ﬁrst person shooter games. Cur-

rently the enemies do not react until they ﬁnd player

entering into their areas. This reactive approach of

enemies often results in repetitiveness of same actions

by the player as well as enemies as they know that all

they can to react to one another’s moves. The repet-

itiveness may result in lack of dynamism and unnat-

uralness for a player after few trials. The proactive

nature of non-player characters (NPCs) may enable

enemies to hunt the player dynamically so that there

would be different hunting strategies and they keep

changing as situation or difﬁculty levels change.

In case of the games like Sims, a player creates

a city for Sims to live in or a theme park to visit.

Once the city’s initial stage of build up is completed,

each Sim do different things to make itself happy.

Sims have means to evaluate their state of happiness

along with set of behaviors for doing what they want

to achieve the happiness(Delgado-Mata and Ib

nez-

Mart

ınez, 2008). When the Sims are not directly con-

trolled by the player, the autonomous notion of agents

can be suggested to make them act in more human-

like manner; for example the Sims preferring swim-

ming over going to library would be able to choose

swimming more frequently than going to library. This

human-like decision making feature of doing things

autonomously can make games more realistic and

fun-oriented.

Teamwork has been considered as one of impor-

tant features in many games including Counter-Strike

(CS) (Schreiner, 2003). For example, in CS play-

ers can join teams of terrorists, counter-terrorists or

become spectators. In each team, the players play

in quite well-organized manner with the mission ob-

jective of trying to eliminate the opposing team in

given time limit. Although the idea of playing in

teams can be observed in non-team members as well

but that is most often on irregular and self-serving

basis. Microsoft’s Halo 3 game can be considered

as another example of the games featuring notion of

teamwork (Mott, 2009). In Halo 3 enemies travel

LevelofDetailbasedAIAdaptationforAgentsinVideoGames

183

and act in groups and in case the player hides for a

minute, most of the group members stay behind to

guard their current location, just couple of them get

assigned to search the player. Apart from searching,

Halo 3 uses teamwork in other situations as well, par-

ticularly when enemies use suppressing ﬁre on the

player meanwhile facilitating other allies to ﬁnd a bet-

ter vantage point or to take cover.

The notions teamwork, autonomy, reactivity and

pro-activeness can be considered as some of most im-

portant features which can be exploited to increase re-

alism and fun-orientatedness in video games. There

has been quite a few video games which incorporate

some of these features. Making all these features ac-

cessible to NPCs requires the kind of abstraction as

provided by Multi-agent systems.

2.2 Multi-Agent Systems

Multi-agent systems can be deﬁned as an agglom-

erations or artiﬁcial societies of “agents”; however

it is quite difﬁcult to precisely deﬁne agents due to

wide differences in agent researchers on the deﬁni-

tion. Wooldridge et al. (Wooldridge and Jennings,

1995) characterizes agents through some well recog-

nized common features of autonomy, reactivity, pro-

activeness and, sociability, which we discuss as un-

der:

1. Autonomy. Autonomy is mean that agents are ca-

pable of acting independently on their own, with-

out any external inﬂuence. This is one of most

important and distinguishing features of agents by

which they exhibit their control over their internal

state.

2. Reactivity. Reactivity implies that agents respond

to the changes occurring in the environment in due

time. In other words for some stimulus there will

be corresponding response. This feature of agents

make them aware as well as responsible to the

environment that which behavior they can adopt

to fulﬁll their design objectives in some particular

situation.

3. Pro-activeness. Pro-activeness makes agents to

behave actively instead of passively. To fulﬁll

their design objectives agents identify the oppor-

tunities and do subsequent actions for achieving

their respective purposes. Pro-activeness enables

agents for making initiatives to achieve goals in-

stead of just reactively responding events in the

environment.

4. Sociability. It is the ability of an agent by which

it interacts with other agents or humans in the

system. Communication is the means to coordi-

nate agent actions, there can be different means of

communication among agents.

2.3 Answering the Challenge of Player

Experience

Player’s interaction with a game can be considered

as one of critical performance metrics for determin-

ing quality game experience. The level of interaction

can be evaluated through the number of frames dis-

played in a second. It is estimated that for interactive

computer graphics, a new picture is to be rendered be-

tween 25 to 30 times in a second. Hence the process

of refreshing a frame need to be done in 30 millisec-

onds or so. The time available to a frame is divided

between rendering and the AI parts of the game. In

the time, the simulation of whole game world requires

a real-time behavior generation of game entities and

their rendering.

A signiﬁcant increase in the computational re-

quirements of a frame would cause delay or lag in

the game and make interaction difﬁcult for the player.

The lag in the interaction may degrade frame rate and

would eventually disrupt the players’ ability to syn-

chronize their actions with the game. Ideally for a

player the interactivity would mean that there is no lag

time in game response. Cumulative game lag can be

said as a sum total of following three factors of lags

namely player input, rendering and AI. Cumulative

game lag can be represented as linear combination of

all three factors in following manner:

Cumulative game lag = Player input

delay + Graphical rendering delay +

Artificial intelligence delay

Although any one of above factors can be sufﬁ-

cient to deteriorate game ﬂow and interactivity. In

this paper we are focused on the AI factor. The pri-

mary motivations for the selecting AI is due to the fact

that among three factors the AI factor can be said as

one of most ignored areas in current research. Even

in the case of time distribution normally game AI gets

signiﬁcantly short time than rendering. According to

Niederberger et al., normally between 5 and 10 per-

cent of the overall CPU time is reserved for AI in a

real-time game (Niederberger and Gross, 2002).

Here our focus on minimizing lag value in game

AI would ensure availability of sufﬁcient computa-

tional resources required for simulating game enti-

ties and above all averting undesirable side-effects on

game AI’s part. To minimize lag in AI part of video

games one of very ﬁrst steps is to identify the factors

which can inﬂuence frame rate.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

184

2.3.1 Game Frame Rate

Usually a frame gets around 30 milliseconds as

“frame time” for both rendering and updating game

AI from the game engine. The frame time for updat-

ing game AI is used for computing and simulating the

AI behaviors of the NPCs in a frame.

The elements inﬂuencing ”frame time” within the

scope of game AI mainly include computational re-

quirements of NPCs and their quantity in a frame.

The frame simulating large number of NPCs or more

complex behaviors would require more than expected

computational resources for game AI. These elements

of video game performance can be considered as the

deciding factors in generation of frames from game

AI’s perspective. In such situation a game may result

in a lower frame rate as game entities would take more

time for simulating agents which require signiﬁcant

computational resources. On the other hand, if the

same agents run on another platform (with increased

resources), it is likely to maintain required frame rate.

In other words, any changes in the computational re-

quirements of NPCs and their quantity would neces-

sarily bring corresponding adjustments in the frame

rate. Ignoring changing requirement of agents would

lead to undesired consequences of either irrelevant AI

or lag into it. In case of more than required compu-

tational resources, game AI can be programmed to

improve its realism and naturalness by introducing

additional features into it. Using additional features

of game AI can prevent unnecessary wastage of re-

sources.

Ideally game engines need to provide some scal-

able and viable means which could provide QoS sup-

port in case there is a change in the performance re-

quirement of agents. Absence of such QoS support

methods may lead to lag in updating game AI and

subsequently reduced frame rate which would even-

tually result in lack of interest and frustration for a

game player.

2.3.2 Bidirectional Adaptation of Agents’

Behavioral ”Complexity” as a Means

of QoS Support

In order to avoid degraded frame rate or resource

wastage we need a ﬂexible mechanism which can

handle performance requirements of NPCs at runtime.

A bidirectional adaptation mechanism can serve the

purpose. The adaptation need to lessen its perfor-

mance requirements when it notices a decrease in the

frame rate and allows requirements in case of other-

wise. This decrease and increase in the performance

requirements can be made possible through the con-

ditional permission to execute certain features in one

case while their prohibition in the other. Here QoS

support would be about suggesting a mechanism for

graceful degradation and progressive enhancement of

agent behaviors.

Graceful degradation is meant to provide an alter-

native version of the game AI in case the system gets

overloaded. In our context, it would take up frame

rate as higher as possible. Progressive enhancement

can be considered as the other side of the QoS coin

in video games. Progressive enhancement works the

other way around of graceful degradation. It starts

with a baseline version of game AI and then fea-

ture enhancement goes till the most sophisticated ver-

sion by maintaining frame rate throughout the pro-

cess. Progressive enhancement can work in the situa-

tions where despite sufﬁcient number of frames there

is no change in the game AI, here it can progressively

increase game AI features.

In our context of game AI in video games, both

graceful degradation and progressive enhancement

can be applied to maintain QoS in most of the situ-

ations. The approach would reserve advanced game

features of game AI and/or gameplay subject to avail-

ability of sufﬁcient computational resources. One of

main differences between graceful degradation and

progressive enhancement lies in determining the point

where each one begins and here the LoD technique

can provide relatively automatic and easy solution for

its determination. The LoD technique also provides

ways by which game agents can have different repre-

sentations.

3 LEVEL OF DETAIL

Level of detail (LoD) technique suggests to modulate

the complexity of a 3D object representation accord-

ing to the distance from it is viewed or any other cri-

teria (Luebke et al., 2002). The technique, as intro-

duced by James Clark (Clark, 1976), is meant to man-

age processing load on graphics pipeline while deliv-

ering an acceptable quality of images. He suggested

that structuring the rendering details can optimize the

processing quality if a 3D object’s visible quality and

details are made to have a correspondence with the

distance from it is viewed.

In ﬁgure 1, we can observe the rational of Clark,

here with changing the number of vertices we see the

change in quality and visualization of a sphere. In

other words, LoD technique trades spatial ﬁdelity for

temporal ﬁdelity in a way that a less complex model

would be instantly rendered by compromising over its

coarser representation.

LevelofDetailbasedAIAdaptationforAgentsinVideoGames

185

Figure 1: Basic concept of LoD: An object is simpliﬁed by

different representations through varying number of poly-

gons

citeluebke2001perceptually.

3.1 How LoD Technique Works?

Since geometric datasets are usually too large in data

size and complex in terms of time and computational

resource demands so their rendering can become a

tedious and time consuming process. The LoD ap-

proach suggests different representations of a 3D ob-

ject model by varying in the details and geometrical

complexity. The geometrical complexity and time de-

mands of a graphical object comes can be determined

in terms of the number of polygons used for its ren-

dering. Usually an object model having more number

of polygons consumes more time and resources than

the one having less number of polygons in its render-

ing. Although there can be other factors involved in

the complexity and resource demand of a graphical

model of an object but the relations between polyg-

onal quantity and resource consumption are gener-

ally considered as established ones (Deering, 1993).

For example, one can clearly determine the difference

in rendering quality by observing following ﬁgure 1

and can draw general conclusion that how number of

polygons affect the rendering quality of an image’s

graphical representational model.

Once these different representations of a model

are on hand (as shown in the above ﬁgure for our ex-

ample), the LoD technique will suggest their selection

at a particular time point based on particular positive

selection bias. The latter can be their size, camera dis-

tance or any other criteria. The net beneﬁt of the ap-

proach would be that only necessary objects that will

get maximum amount of processor time and resources

at one time point while others are either shown with

less details or just ignored at that time point. These

objects can get less shared resources when the situ-

ation changes and their importance is substituted by

other objects. Basically, the approach is inspired by

rendering of the polygonal geometry of object models

by using coarser LoDs for distant or invisible objects

(David Luebke, 2003).

3.2 LoD in Game AI

The LoD technique can be used in game AI for se-

lecting most interesting agents and correspondingly

priorirtizing them for computational resources distri-

bution. The priority in computational resource dis-

tribution would ensure that the selected agents would

have access to the game AI which requires more com-

putational resources. The agents having lower priority

value would get the share in computational resources

required for a basic version of the game AI. In other

words, the technique focuses on ﬁnding relatively in-

teresting or irrelevant characters in a scene and pro-

gressively upgrade or turn down their behavioral de-

tails. Evidently, the notion of LoD in game AI may

not be regarded as an exact transfer of the technique

from rendering part as the nature and issues partic-

ular to the game AI part would be different, despite

the fact that basic concept remains the same. Here we

need to address some speciﬁc issues which are partic-

ularly related to the AI part of the games.

Objects varying in priorities and then in simulat-

ing most relevant ones is common to both render-

ing and AI based LoD, the primary difﬁculty lies in

the notion of importance and its different representa-

tions. The notion of distance or visibility may not be

in the same form or as clear as it is in the rendering

part. Neiderberger et al. (Niederberger and Gross,

2005) argue that invisible objects do not stop living

even when they are out of visible area. Moreover, in

some situations, even the objects which are only par-

tially visible in a scenario may need more detailed AI

compared with the ones which are currently visible.

Hence the distance from the camera matters as much

as the behavioral functionality in a scenario. Here the

notion of importance may not be resolved solely by

the camera distance. In addition to that we need to

address the problem of how to provide different rep-

resentations of an agent.

Hence, to deal with this problem, we suggest us-

ing agent organizations as a means to provide differ-

ent representations of agents and decide their impor-

tance with respect to other agents in the game AI.

4 RELATED WORKS

In 2002, Brockington (Brockington, 2002) presented

a paper on extending the usage of existing computer

graphics based LoD to game AI. Brockington uses

LoD for the Neverwinter Nights(NWN) game. He di-

vides NWN game creatures under ﬁve classiﬁcations

for determining their update frequencies. The classiﬁ-

cation determines a creature’s LoD and subsequently

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

186

percentage of CPU time for each LoD. The approach

argues the interest of dedicating most of CPU time

(60%, to be precise) for computing best algorithms

to Player characters (PCs) as they always remain on

the screen. The CPU time percentage and priority get

decreased with increased LoD value according to the

classiﬁcation of game creatures.

Wißner et al. (Wißner et al., 2010) have tried to

generalize the approach by extending LoD technique

to the navigation, movement updates, collision avoid-

ance and behavior execution of game entities based

on 8 LoD levels. The approach’s signiﬁcant fea-

tures include the diversity and heterogeneity of be-

haviors which can be reduced with it. Moreover,

the behavior simpliﬁcation is applied at the behav-

ior execution stage rather than during the behavior

selection process. Another signiﬁcant feature of the

approach is the addition of visibility factor besides

traditional camera distance based LoD implementa-

tions. Explaining visibility criterion besides camera

distance, the authors argue that some situations may

make agents relevant and important enough for show-

ing their full behavioral details regardless of their

camera position.

Kistler et al. (Kistler et al., 2010) extend the work

by (Wißner et al., 2010) using the same parameters.

The only difference between these works is the Kistler

et al. ’s detailed explanation of their implementa-

tion of the approach on Virtual Beer Garden while

Wißner et al. (Wißner et al., 2010) are more inter-

ested in Augsburg3D along with addition of two value

of LoDs making it a total of 10 LoDs which were 8 in

the earlier case.

Osborne et al. (Osborne and Dickinson, 2010) in-

troduce a hierarchical approach for presenting agent

behavior details according to the camera distance.

The main problem addressed by this work is the sim-

ulation of large number of agents acting in groups.

Neiderberger et al. (Niederberger and Gross,

2005) ’s work can be considered as one of the most

comprehensive works in the domain. This work not

only applies LoD on both individual and collective

behavior levels, but also provides a number of 21

LoDs, which is the highest number we encountered in

the literature. They suggest navigation, path planning

and collision avoidance as AI behaviors of game enti-

ties. LoD is also applied for scheduling of agents, col-

lision avoidance, path planning and group decisions

of agents. The LoD approach is implemented on the

combined distance and visibility factors for effective

selection of entities. A special scheduling algorithm

distributes simulation time to the agents depending on

their visibility and camera distance.

5 LoD FRAMEWORK

The framework proposed uses the LoD technique to

provide different levels of behaviors for an agent.

Each level varies in computational overhead and be-

havioral complexity. To address the challenge of

frame rate, the best possible level is allowed to carry

out its tasks while the more complex behaviors are

subject to availability of additional computational re-

sources. When the frame rate degrades, a behavior

requiring less computational resources is permitted.

Besides, to assign priorities amongs agent behaviors,

we use an organizational model so that priorities are

related to roles.

5.1 Organizational Model of Agents

Our LoD framework uses agent organization to ad-

dress two issues:

• Organizing and representing the different behav-

iors of an agent so that, whenever there is a change

in the frame rate, a particular representation of an

agent is always available.

• Evaluating the relative importance of an agent in

a particular situation in order to select the most

appropriate agents’ representations. That is, when

the system is overloaded, the organization settings

are used to decide which agents can be abandoned

altogether since they are less important than oth-

ers in the situation.

So we follow the AGR model (Ferber et al., 2003) for

the organizational structure of agents. Here we brieﬂy

deﬁne the AGR model. The AGR (Agent, Group,

Role) model advocates organization as a primary con-

cept for building large scale and heterogeneous sys-

tems. The model does not focus on the internal archi-

tecture nor the behavior of individual agents but sug-

gests organization as a structural relationship between

collection of agents. The AGR model deﬁnes three

main concepts as its basis for an organizational struc-

ture: agent, as an active and communicating entity;

groups are comprised of agents in the set by tagging

them under a collection; ﬁnally an agent’s functional

representation in a group is given by deﬁning its role.

The AGR model helps us to deﬁne how different

behaviors can be glued together to construct an agent

role in an agent society. Different combinations of be-

haviors would create distinct roles which would serve

as representations of an agent. The behavioral bond-

ing provided by AGR allows us to have different rep-

resentations of an agent through multiple roles. The

model provides an organizational methodology which

can be used to program agents which can have differ-

LevelofDetailbasedAIAdaptationforAgentsinVideoGames

187

ent representations to act within a justiﬁable time ac-

cording to the available computational resources.

The way AGR associates different roles in a group

makes it convenient to deﬁne importance of a role in

a particular situation. Here agent roles can be as-

sociated together according to their functional roles

or/and availability of computational resources in a

game environment. Taking an example, when a group

of agents moves to a particular point, the path ﬁnd-

ing result of each agent would be almost the same.

This repeatability of the path ﬁnding algorithm can

be avoided, if we can glue all agents in the collec-

tion in some organizational structure. Here our or-

ganizational approach can guide us to group differ-

ent game entities according to their roles. In this ex-

ample of path ﬁnding, some speciﬁc agents having

group leader role would be allowed to access particu-

lar behavior while others agents would just follow the

group leader.

Organization is also used for quantifying and ﬁl-

tering agent roles in a frame instead of normally used

distance or size measures. A programmer speciﬁes

conﬁguration ﬁle lists deﬁning the allocated compu-

tational resources for the agent roles according to our

model of organization. The game engine uses the con-

ﬁguration ﬁle based for distributing the given com-

putational resources to agents. The game engine es-

timates an elementary value for the frame rate as a

starting point and then uses it as a metric of shifting

the QoS support modes for the agents AI.

5.2 QoS Support Modes

An agent requires a certain number of computational

resource units for carrying out particular set of actions

in a particular role. These resource requirements of an

agent are variable and depend on the computations in-

volved in a role. If a handful agent play roles which

require realtively more computational resources than

affordable by the game engine it would necessarily re-

sult in degraded frame rate. On the other hand, if the

game engine can afford to provide sufﬁcient compu-

tational resources it would be better to use them for

providing higher levels of game AI rather than simply

wasting them.

A QoS support mode associates the frame rate

with the computational resources distribution of game

AI agents. The approach provides several levels of

QoS support by encompassing multiple modes of

game AI. The game engine deﬁnes a threshold level

of frame time for each mode and any noticable change

in the frame time would switch the mode and subse-

quently computational requirements of game AI. In

other words, a QoS support mode governs the policy

by which computational resources are distributed to

the game AI agents in a frame. The game engine uses

multiple QoS support modes for determining which

agents would be allowed to play a speciﬁc role at

a particular time. The approach provides QoS sup-

port by adding ﬂexibility and differentiation levels for

game engine according to the available computational

resources for a frame.

The available resource units for an agent are de-

termined by the mode in which it is operating. Once

an agent gets resource units awarded, it can access a

particular role in the organisational based agent soci-

ety. An increase or decrease in frame time switches

the particular mode on or off. So that depending on

the mode there is a difference in the QoS support and

role access for the agents. This way, the modes ap-

proach will not fumble optimal utilization of com-

putational resources with increasing demand of re-

sources as there would be QoS support for all possible

situations in form of relevant modes. By switching

QoS modes, our approach provides programmers the

ﬂexibility and versatility to meet various time perfor-

mance needs for the agents.

5.3 Agent Life Cycle Model:

Perception-Deliberation-Action

Model

Normally, an agent model deﬁnes a set of activites

and the data and control ﬂow among them as a gen-

eral methodology for carrying out any task. The set

of activities and their interactions would deﬁne the

life cycle of an agent. A brief description of agent

activities is given as under:

1. Perception. It is about observing and getting a

sensing stimulus from the game environment. On

receiving a stimulus, the perception mechanism

decides further meaning of the stimulus. More-

over, the perception mechanism examines and de-

termines whether to take into account the sensing

stimulus as an input or ignore it if found irrele-

vant. Assigning computational resources in terms

of tokens would allow developers to specify the

time taken by agents for perceiving their environ-

ment.

2. Deliberation. This activity can be considered as

the central process in the agent life cycle. The

main purpose of the activity is to determine the

next action of an agent. An agent deliberates ac-

cording to its perception of the environment and

its deliberation model. Depending upon the con-

text and requirements, an agent’s deliberation can

be very simple like stimulus-response activity for

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

188

reactive agents to very complex ones using spatial

reasoning or any other algorithms. Because our

game agent model is independent of any particu-

lar deliberation approach, game programmers will

be able to implement any deliberation methodol-

ogy.

3. Action. The action step provides means for an

agent to select and carry out an action in its envi-

ronment. Considering possible outcomes of the

deliberation step, an agent selects one or more

actions to perform in its environment. Once an

agent selects an action, it is executed by updat-

ing its state or the environment. Depending on the

available computational resources, an action is as-

signed a speciﬁc number of tokens. The tokens

can be considered as the computational resources

required for changing the agent’s state or making

the effects of the action appear in the environment.

Figure 2: Agent life cycle.

Agent life cycle is a dynamic, iterative process of

input handling, incorporating deliberation mechanism

and carrying out actions in the agent environment.

The agents would need computational resources to

carry out a particular activity. The agents need to

consider the computational requirements involved for

performing any task. The life cycle of an agent can be

summarized in ﬁgure 2. The ﬁgure describes the life

cycle of an agent as it evolves and shows the process

that the agent description supports for input, infer-

ence, and output modules in the agent environment.

5.4 Valued Sensing and Action Model

for Game Agents

Game agents use different sensing and actions to carry

out their tasks corresponding to their roles in the or-

ganization. Each sensing and action task requires

some computational resources for its execution. In

other words, behavioral tasks come with certain cost

in terms of resources. The execution cost of a task

varies according to the computations involved in it.

For example, a random move task could has a low

cost compared to a path ﬁnding task requiring more

CPU cycles. So, agent behaviors come with an abso-

lute cost in terms of CPU cycles.

Distributing computational resources with an ab-

solute measurement of agent behaviors would make

it practically impossible to program game agents as

differences in resources for different game platforms

would require customized programming for every sin-

gle hardware conﬁguration. Moreover, the absolute

cost measurement of behaviors would not help in

differentiating among different versions of behaviors

and later selecting a cost effective one as it would

require comparisons of memory and CPU usage be-

tween different behaviors. Therefore we propose “re-

source tokens” as an abstract notion of computational

resources.

5.4.1 Tokens

Here in our proposition, we use the logical notion of

computational resources in our framework. For exam-

ple, the logical time associates the wall clock time and

the QoS modes. The wall clock time measures the ex-

ecution time of a task on physical clock, for example a

move action may require 40 milliseconds on a speciﬁc

processor, while a path ﬁnding action takes 120 mil-

liseconds on the same processor. A correspondence

between the logical and physical notion of computa-

tional resources would make it their distribution an

easier job for the programmers.

A “token” can be described as an abstract mea-

surement unit for the notion of computational re-

sources for weighing agent behaviors. The idea relies

on providing an independent behavior measurement

abstraction which is only related to the complexity of

the behaviors, and not to the hardware conﬁguration.

Each sensing and action task is assigned a value in

terms of tokens through a conﬁguration ﬁle. So, the

programmer relatively analyses and evaluates each

task with respect to other tasks and bounds their ex-

ecution according to a determined number of tokens.

The game engine evaluates which agents can be re-

quested in a particular game loop iteration and assigns

them tokens by reading the conﬁguration ﬁle. The

token assignment is based on the computational re-

sources based relative evaluation of agent tasks. This

relative value is assigned as per the absolute compu-

tational cost of different tasks in the game. For exam-

ple, a move action might take 2 tokens of logical time

and path ﬁnding action takes 6 tokens and after exe-

cuting these two actions 160 milliseconds of physical

time might have passed.

5.5 Agent Delegation Model

Delegation can be explained as a request of one en-

tity to another for performing some actions on behalf

of the former. Our approach generates and processes

actions by delegating them to the game engine as ser-

vices. The action delegation model is based on the

LevelofDetailbasedAIAdaptationforAgentsinVideoGames

189

notion of providing standard and consistent mecha-

nisms to generate and process actions independently

from any speciﬁc hardware conﬁgurations. Basically,

our delegation model for agents provides a resource

request mechanism for access to the sensing and ac-

tion services from the game engine. The life cycle of

an agent using delegation can be summarized in fol-

lowing ﬁgure 3.

Figure 3: Agent life cycle using delegation.

In fact, the interest of the delegation model relies

on the idea that in order to limit the incompatibility of

computational resources among different game plat-

forms, we need another entity for carrying out tasks

delegated by agents. In our case, the game engine

plays that role and, ensures availability of speciﬁc ser-

vices and their token-based distribution to the game

agents. So, an agent can request for delegating the

sensing and action tasks to the game engine according

to its available number of tokens. But thanks to del-

egation, when computing resources are lacking, the

game engine can now answer with a simpliﬁed ver-

sion of the same service requiring less computations

(i.e. tokens) by using our LoD approach. This request

and grant model of services delegation in the context

of LoD mainly requires two things:

1. Tokens. Tokens are relative measurement abstrac-

tion units which are used to differentiate between

different costs of behaviors.

2. Services Library. A services library is main-

tained by the game engine for assembling all ser-

vices and their corresponding cost in terms of to-

kens. In this way, once an agent submits a request,

the game engine can easily search and grant or

refuse the service according to the cost of the ser-

vice and available tokens to the agent.

In our service delegation model, the services must

associate with a source role for receiving a noti-

ﬁcation. This service and role bonding associa-

tion provides an important beneﬁt: request notiﬁ-

cations are sent to only those services that want to

receive them for example, decoration clouds can

not send pathﬁnding service request. The main

beneﬁt of this design is that the perception and

action AI that processes these services is cleanly

separated from the deliberation logic which gen-

erates these requests. This separation of percep-

tion/action and deliberation logic serves our pur-

poses of delegating a game agent’s processing

through a separate piece of code namely a service.

Agent actions are decomposed into small inde-

pendent services. The approach ensures to carry

out the separation of concerns in different hard-

ware conﬁgurations. The game engine maintains

a collection of services and correspondingly re-

quired tokens for the sensing and action services.

As the services library would be used for selecting

appropriate services according to the LoD level so

services having different simpliﬁed and advanced

versions are provided in the library; however their

degree of resource requirements is associated with

the number of tokens for them. The services re-

quiring more computational requirements would

have relative higher cost so they would require

comparatively large number of tokens and the

simpliﬁed ones (requiring lesser CPU time) would

require relatively less number of tokens. In this

way, when a game engine receives a requests for a

particular sensing or action service from an agent,

it looks up the services library. On the availabil-

ity of the service, If the agent qualiﬁes for the ser-

vice in its functional role and has required number

of available tokens, the service would be granted

otherwise it will be refused.

6 THE PILOT EXPERIMENT

The objective of this experiment is to evaluate the

impact of using the LoD based AI adaptation on a

player’s interactivity and game experience.

6.1 The Game: My Duck Hunt

My Duck Hunt is a single player shooting game

developed with AGDE framework http://gforge-

lirmm.lirmm.fr/gf/project/agde/frs. The player con-

trols a reticle on the screen in order to acheive follow-

ing goals :

• Kill ducks before they leave the scene.

• Avoid to kill ﬂamingo.

• Prevents gombas from eating ﬂamingos.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

190

Figure 4: A screenshot of My Duck hunt.

Figure 5: Pictures of the experiment.

Clouds evolve in the sky as decoration elements. The

ﬁgure 4 shows a screen shot of the game.

6.2 LoD Conﬁguration

In this experiment, we use three modes: advanced,

best effort and basic for distributing resource tokens

among agents. One of following three modes get

switched depending on the available time for a frame.

1. Advanced Mode. This mode of QoS supports

full features of agents and it is triggered when the

time for a frame exceeds the estimated range of

time (i.e FrameTime >t1). The advanced mode

of game AI operates, hence agents have access to

most of the behaviors with sufﬁcient resource to-

kens that are required for their execution. In this

mode, entities have the access to the full set of

actions that have been deﬁned.

2. Best Effort Mode. The normalized mode gets

triggered when time for a frame lies in the esti-

mated range ( i.e. t0 <FrameTime <t1). In this

case, only most important behaviors of agents that

get resource tokens. In this mode clouds can not

move.

3. Basic Mode. This degraded mode provides ba-

sic functionality of game agents despite frame rate

falls behind the required range ( i.e. FrameTime

<t0). The basic design of game AI remains ac-

cessible to agents instead of total collapse or lag

in the game AI. In this mode, gombas can not per-

form the Jump action.

In order to simulate a decrease of available resources,

an entity can be added to the game. This entity send

an action that requires important amount of resources

in Optimal and Medium modes.

6.3 Participants

The prototype test was conducted on 8 subjects hav-

ing ages between 23 and 28 years old. Most of

the participants reported themselves as regulars video

game players playing atleast once a week except two

candidates that do not play video games.

6.4 Protocol

The experiment follows a repeated-measures design.

Subjects have to play the Duck Hant game on the

same computer. Two versions of the game are avail-

able, one that includes LoD based AI adaptation and

the other without AI adaptation. The experiment pro-

ceeds as follows:

1. The candidate gets a quick introduction about the

game. This tutorial aims to provide an introduc-

tion to game’s objectives and how controlling the

game.

2. The candidate plays the two versions of the game

with and without LoD ( the order is random).

Each game version consists of eight stages (called

waves). The subjects are not informed about the

difference between the two sessions. Throughout

the session, the subjects were asked to report their

feelings of ”lag” by pressing the space button on

the keyboard and we note the time of the button

press and the total time of the wave. The time ra-

tio of ”space button” to the total time of a vague

clearly signiﬁes the participants’ preference of our

approach over the one not using it.

3. At the end of each wave the subject has to evaluate

the quality of the interaction during each waves.

4. A questionnaire is proposed and the subjects were

interviewed about the quality of the interaction

during both versions of the game and about his

opinion concerning the game experience.

The ﬁgure 5 shows two pictures of the experiment.

The ﬁrst on the left shows a subject who evaluates

LevelofDetailbasedAIAdaptationforAgentsinVideoGames

191

the interaction after a game wave. The second on

the right shows a subject plying the game.

6.5 Hypothesis

The experiment aims to demonstrate that the pro-

posed LoD based adaptation inﬂuences the game ex-

perience. Our main objective is to provide a game

experience in which (i) game AI can be adapted to

the requirements of the platform; (ii) the adaptation is

meant to provide an increased level of user experience

which is not the case otherwise. To achieve above

mentioned objectives, we state following hypothesis:

• H.A.0: There is no difference of interactivity and

game experience between the game version with

LoD adaptation and the other without LoD adap-

tation.

• H.B.0: There is no difference concerning the

global perceived quality of interaction between

the game with LoD and game without LoD.

• H.C.0: There is no difference of the ratio of time

of holding the key space and the session duration

between the game with LoD and game without

LoD.

6.6 Results

Statistical analysis was performed using

R(http://www.r-project.org/) version 2.15.0. We

have used t-paired tests to reject the three hypotheses.

The results of the t-tests on the eight waves are

summarized in table 1. The t-tests have rejected

the ﬁrst hypothesis H.A.0 in the ﬁrst six waves.

The difference between the perceived quality of

interaction using the game with LoD and without

LoD was statically signiﬁcant with a df=7. As for

wave 6, 7 and 8 the hypothesis has not been rejected

with the t-test.

Table 1: Results.

Wave M t(7) p-value

Wave 1 -1.75 -7 0.0002116

Wave 2 -2.25 -13.7477 2.541e-06

Wave 3 -2.25 -13.7477 2.541e-06

Wave 4 -2 -7.4833 0.0001392

Wave 5 -1.625 -5.017 0.001536

Wave 6 -0.25 -0.5092 0.6263

Wave 7 -0.125 -0.2047 0.8436

Wave 8 0.375 2.0494 0.0796

The second hypothesis H.B.0 has been rejected by

t-paired test, thus the difference between the global

perceived quality of interaction of game with LoD and

without LoD was statically signiﬁcant with a mean of

the differences M=-0.75 , t(7) = -2.3932 and a p-value

= 0.04794.

The ratio of session duration and time of hitting

the key space during the game with LoD and without

LoD was different as the hypothesis H.C.0 was re-

jected using t-test. The difference was statically sig-

niﬁcant with a mean M= 0.2763045, t(7) = 9.5683 and

p-value = 2.86e-05.

6.7 Discussion

The results reported by our experiments clearly show

the effect of LoD adaptation on the participants’ game

experience. The ﬁrst hypothesis H.A.0 has been re-

jected by the t-test in six waves. This mean that the

players have perceived a signiﬁcant impact of LoD

adaptation on the interactivity and game experience.

The experimental results signify the the link between

the participants’ game experience and the frame rate

based QoS support for game AI. The game experience

and QoS link can justiy the participants bidirectional

adaptation of game AI or a non-ﬂexible AI. The ar-

gument has supported by the questionaire when the

participants were asked the question: “whether they

would perfer having a ﬂexible AI to maintain inter-

activity or not?”, 6 among 8 participants replied afﬁr-

matively, one of them supported it on the type of the

game while one participant perfered to quit the game

if quality game expereince is visibly compromised for

interactivity or the reverse.

However, the hypothesis has not been rejected for

the last three waves as no signiﬁcant difference has

been reported in the experiments. This can be ex-

plained by the fact that starting from the sixth wave

the average FPS between the game without LoD and

with LoD is almost the same. For example the ﬁg-

ures 6 and 7 show the average FPS on the ten later

seconds for a candidate and for each playing session.

The last three waves start at 3:45 on the game without

LoD and 3:31 one the game with LoD these moments

are shown on the two ﬁgures by the red lines). We

can see that the average FPS is up to 60 FPS for the

game without Lod (ﬁgure 6) where the average FPS

is around 59 FPS on the game with LoD (ﬁgure 7).

The second hypothesis has been rejected as the

participants report a signiﬁcant degree of difference

between the perception of the quality of interaction

between the game with LoD and without it. The re-

sults support our approach for improving the overall

quality of interactivity of the game experience. The

statical results show that adapting the game AI as per

the computational requirements of game agents sig-

niﬁcantly increased the overall interaction rating from

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

192

Figure 6: Average FPS per second without LoD adaptation

of participant 2.

Figure 7: Average FPS per second with LoD adaptation of

participant 2.

the participants.

Finally, the t-test results have also rejected the hy-

pothesis H.C.0. Here The experiment showed that the

participants playing the without LoD version of the

game reported signiﬁcant difference as compared to

the with LoD version.

7 CONCLUSIONS

In this paper, we have proposed a LoD based adap-

tation technique to prevent resource costly behaviors

of game agents from degrading the quality of interac-

tion. The main idea of our proposition is to consider

the game as an agent organization where each agent

has a role to play meanwhile roles are prioritized by

the game designer. We introduce the logical notion

of resource by which a programmer can associate the

priority of roles to the distribution of resources to the

agent roles. The logical notion of resources make

game agents to delegate their tasks to the game en-

gine and in this way agents’ direct manipulation of

resources is avoided. The delegation model eventu-

ally results in the better frame rate as agents can only

consume the assigned resources.

We tested the effects of LoD in game AI with sev-

eral players in our prototype. We statistically prove

that there is signiﬁcant difference in player’s interac-

tion and game experience when approach is used. As

a future work, we plan to extend our agents to coordi-

nate model to include more found that the technique

signiﬁcantly improved performance, and that players

did not rate improvements in lag as more difﬁcult or

frustrating to use. Our study improves understanding

of LoD in game AI and of how to link it for improving

interaction in video games.

REFERENCES

BioWare (2002). Neverwinter nights - game information.

[Online; accessed 04-September-2012].

Brockington, M. (2002). Level-of-detail ai for a large

role-playing game. AI Game Programming Wisdom,

1:419–425.

Charles, D. (2007). Biologically inspired artiﬁcial intelli-

gence for computer games. Information Science Ref-

erence.

Clark, J. (1976). Hierarchical geometric models for visi-

ble surface algorithms. Communications of the ACM,

19(10):547–554.

Clark, J. (2008). Alief and belief. Journal of Philosophy,

105(10):634–663.

David Luebke, Martin Reddy, J. D. C. A. V. B. W. R. H.

(2003). Level of detail for 3D graphics. Morgan Kauf-

mann Pub.

Dechesne, F., Hattori, H., ter Mors, A., Such, J. M.,

Weyns, D., and Dignum, F., editors (2012). Advanced

Agent Technology - AAMAS 2011 Workshops, AM-

PLE, AOSE, ARMS, DOCM3AS, ITMAS, Taipei, Tai-

wan, May 2-6, 2011. Revised Selected Papers, volume

7068 of Lecture Notes in Computer Science. Springer.

Deering, M. (1993). Data complexity for virtual reality:

where do all the triangles go? In Virtual Reality

Annual International Symposium, 1993., 1993 IEEE,

pages 357–363. IEEE.

Delgado-Mata, C. and Ib

nez-Mart

ınez, J. (2008). Ai

opponents with personality traits in

uberpong. In

Proceedings of the 2nd international conference on

INtelligent TEchnologies for interactive enterTAIN-

ment, page 1. ICST (Institute for Computer Sci-

ences, Social-Informatics and Telecommunications

Engineering).

Dignum, F., editor (2011). Agents for Games and Simu-

lations II - Trends in Techniques, Concepts and De-

sign [AGS 2010, The Second International Workshop

on Agents for Games and Simulations, May 10, 2010,

Toronto, Canada], volume 6525 of Lecture Notes in

Computer Science. Springer.

Dignum, F., Bradshaw, J. M., Silverman, B. G., and van

Doesburg, W. A., editors (2009a). Agents for Games

and Simulations, Trends in Techniques, Concepts and

Design [AGS 2009, The First International Workshop

LevelofDetailbasedAIAdaptationforAgentsinVideoGames

193

on Agents for Games and Simulations, May 11, 2009,

Budapest, Hungary], volume 5920 of Lecture Notes

in Computer Science. Springer.

Dignum, F., Westra, J., Van Doesburg, W., and Harbers,

M. (2009b). Games and agents: Designing intelligent

gameplay. International Journal of Computer Games

Technology, 2009:1–18.

Ferber, J., Gutknecht, O., and Michel, F. (2003). From

agents to organizations: An organizational view of

multi-agent systems. In AOSE, pages 214–230.

Kistler, F., Wißner, M., and Andr

e, E. (2010). Level of

detail based behavior control for virtual characters. In

Intelligent Virtual Agents, pages 118–124. Springer.

Luebke, D., Watson, B., Cohen, J. D., Reddy, M., and

Varshney, A. (2002). Level of Detail for 3D Graph-

ics. Elsevier Science Inc., New York, NY, USA.

Millington, I. and Funge, J. (2009). Artiﬁcial intelligence

for games. Morgan Kaufmann.

Mott, K. (2009). Evolution of artiﬁcial intelligence in

video games : A surve. In Term Papers prepared for

810:161, Artiﬁcial Intelligence, Spring 2009 Univer-

sity of Northern Iowa (UNIAI-09).

Niederberger, C. and Gross, M. (2005). Level-of-detail for

cognitive real-time characters. The Visual Computer,

21(3):188–202.

Niederberger, C. and Gross, M. H. (2002). Towards a game

agent.

Niederberger, C. B. (2005). Behavior Modeling and Real-

Time Simulation for Autonomous Agents using Hierar-

chies and Level-of-Detail. PhD thesis, Swiss Federal

Institute of Technology, ETH Zurich.

Orkin, J. (2006). Three states and a plan: the ai of fear. In

Game Developers Conference, volume 2006. Citeseer.

Osborne, D. and Dickinson, P. (2010). Improving games ai

performance using grouped hierarchical level of de-

tail.

Rabin, S. (2002). AI Game Programming Wisdom. Charles

River Media, Inc., Rockland, MA, USA.

Remendy (2001). Games by remedy. [Online; accessed

04-September-2012].

Russell, S. and Norvig, P. (2010). Artiﬁcial intelligence: a

modern approach. Prentice hall.

Schreiner, T. (2003). Artiﬁcial intelligence in game design.

Studios, L. (2001). Black and white. [Online; accessed

04-September-2012].

Valve (2008). Left 4 dead blog. [Online; accessed 04-

September-2012].

Wißner, M., Kistler, F., and Andr

e, E. (2010). Level of de-

tail ai for virtual characters in games and simulation.

Motion in Games, pages 206–217.

Wooldridge, M. and Jennings, N. (1995). Intelligent agents:

Theory and practice. Knowledge engineering review,

10(2):115–152.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

194