Automated Scenario Generation

Coupling Planning Techniques with Smart Objects

Gwen R. Ferdinandus

, Marieke Peeters

1,2

, Karel van den Bosch

and John-Jules Ch. Meyer

Information and Computing Sciences, Utrecht University, Utrecht, The Netherlands

Training & Performance Innovations, TNO, Soesterberg, The Netherlands

Keywords:

Automated Scenario Generation, Adaptive Educational Games, Serious Games, Scenario-based Training.

Abstract:

Serious games allow for adaptive and personalised forms of training; the nature and timing of learning ac-

tivities can be tailored to the trainee’s needs and interests. Autonomous game-based training requires for the

automatic selection of appropriate exercises for an individual trainee. This paper presents a framework for an

automated scenario generation system. The underlying notion is that a learning experience is deﬁned by the

objects and agents that inhabit the training environment. Our system uses automated planning to assess the

behaviour required to achieve the (personalised) training objective. It then generates a scenario by selecting

semantically annotated (or ‘smart’) objects and by assigning goals to the virtual characters. The resulting

situations trigger the trainee to execute the desired behaviour. To test the framework, a prototype has been

developed to train the First Aid treatment of burns. Experienced instructors evaluated scenarios written by

three types of authors: the prototype, ﬁrst-aid experts, and laymen. The prototype produced scenarios that

were at least as good as laymen scenarios. First-aid experts seemed the best scenario writers, although dif-

ferences were not signiﬁcant. It is concluded that combining automated planning, smart objects, and virtual

agent behaviour, is a promising approach to automated scenario generation.

1 INTRODUCTION

Serious games have become increasingly popular as

educational tools. Advances in graphic and AI tech-

niques have provided us with virtual environments,

inhabited by believable characters, where the trainee

can practise the learning tasks autonomously. To en-

hance the effectiveness of these environments, serious

games incorporate features from efﬁcacious training

forms (e.g. Peirce et al. (2008) and Peeters et al.

(2012). A training methodology that, because of its

story-like nature, lends itself especially well for this

prupose is Scenario-based Training (SBT). SBT pri-

marily concentrates on the type of exercises, i.e. con-

textualized, whole-task storylines, exemplifying the

learning-by-doing approach to training (Oser et al.,

1999; Salas et al., 2006). In SBT the trainee is con-

fronted with a representative sequence of events (the

scenario) within a simulated environment (e.g. the

game). Of course, most training methodologies also

recommend a certain ordering of learning tasks (Mer-

rill, 2002). However, the planning and ordering of

learning tasks is not the issue in the work presented

here.

The problem discussed in this paper is that once

such an ordering has been established, suitable exer-

cises need to be created to provide the trainee with

practice opportunities regarding the selected learn-

ing task. Since manual scenario creation is a time-

consuming process, most systems reuse a limited set

of scenarios linked to each learning task, however,

after several occasions, scenario repetitions becomes

inevitable. This is a problem for training directed

at skill maintenance, i.e. continued training. More-

over, to ensure effectiveness, the exercises (scenarios)

should be adapted to the individual needs and abil-

ities of the trainee, and offer him varied experiences

(Peeters et al., 2012). As such, the need for automated

scenario generation arises.

In this paper we propose an automated scenario

generation framework to produce training scenarios

that encompass a previously selected learning objec-

tive while warranting complete and coherent story-

lines. The next section ﬁrst considers related work

on automated scenario generation. Section 3 then de-

tails the design, followed by the evaluation in Section

4. Finally, Section 5 discusses the implications of our

research and opportunities for future work.

R. Ferdinandus G., Peeters M., van den Bosch K. and Ch. Meyer J..

Automated Scenario Generation - Coupling Planning Techniques with Smart Objects.

DOI: 10.5220/0004354600760081

In Proceedings of the 5th International Conference on Computer Supported Education (CSEDU-2013), pages 76-81

ISBN: 978-989-8565-53-2

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

2 RELATED WORK

There is no generally accepted solution to scenario

generation, despite it being a growing subject of in-

terest. This section discarusses some existing ap-

proaches.

The two works discussed below focus on the cre-

ation of a game world that matches the intended learn-

ing task. Martin et al. (2009) propose to construct an

initial scenario based on the training task, which in

turn is extended by adding events to increase the com-

plexity level. The requirements on the game world

that follow from the resulting conceptual scenario are

addressed using a shape grammar. Lopes and Bidarra

(2011a) also focus on the realisation of the scenario

within the virtual world, arguing that the contents of

the game world determine the trainee’s experience.

They propose the use of Smart Objects (Kallmann and

Thalmann, 1998), which are annotated with the ser-

vices they offer to their surroundings, such as the ex-

periences they could offer a player. These annotations

can be used to steer the content selection process. Al-

though in both of these approaches content generation

sprouts from the initial learning objective, the lack of

an explicit task representation makes it difﬁcult to in-

terpret the resulting training scenario within the con-

text of the training domain. It is also impossible to

derive the expected action plan for the trainee.

In contrast, Niehaus and Riedl (2009) employ au-

tomated planning techniques to construct a scenario

based on the trainee’s expected action sequence. By

adapting a default scenario plan, consisting of ordered

high level tasks, these action sequences can be ad-

justed to the needs and abilities of the trainee while

maintaining a coherent storyline . An important ad-

vantage of this approach is the possibility to track the

actions the trainee is required to perform to accom-

plish the learning task. Moreover, the action sequence

has been derived from an explicit representation of the

training domain, which promotes the interpretation of

the scenario with respect to the learning content.

The next two approaches select events to persuade

the trainee into performing the desired actions. Grois

et al. (1998) employ probabilistic networks to com-

pute a set of events likely to cause an opportunity for

practising the learning task. Zook et al. (2012) use a

basic set of events and, subsequently, use a genetic al-

gorithm to extend, mutate and improve the sequence

of events until an acceptable scenario has been gen-

erated. Both approaches offer interesting alternatives

(or additions) yet require very speciﬁc data such as

probability functions and quantitative scenario evalu-

ation functions, that are all but trivial to deﬁne. None

of the last three approaches discuss how the scenario

should be facilitated by the game world.

3 DESIGN

This section describes the Scenario Generator frame-

work. First, we consider the requirements for effec-

tive training scenarios that have driven the design pro-

cess. Then, the general design is presented, followed

by an explanation of the framework’s two main com-

ponents: the Action Planner, and the Object Selector.

3.1 Scenario Requirements

To understand which elements determine the effec-

tiveness of a training scenario we have gathered in-

formation from literature studies (Martin et al., 2009;

Issenberg et al., 2005; Peeters et al., 2012) and in-

terviews with experienced instructors from different

training domains (i.e. First Aid, In-Company Emer-

gency Assistance, and the Dutch Royal Navy). From

this research, we concluded that in order to be ef-

fective a training scenario must be 1) focused on the

learning objective, 2) adapted to the trainee’s com-

petency level, 3) representative of real life situations,

and 4) complete with respect to high level procedures.

The last requirement is illustrated by the following

example: a ﬁrst-aid training exercise should not be

restricted to the treatment of the victim, but each ex-

ercise should also contain the preceding steps of se-

curing the environment and determining the problem,

since real life situations will never encompass merely

the treatment of the victim.

3.2 The Scenario Generator

The Scenario Generator has been designed to work

within the context of an agent-based Adaptive Edu-

cational Game (AEG). Figure 1 depicts the general

design of the Scenario Generator within the context

of the AEG. The system consists of two main parts:

an Action Planner and an Object Selector. The Ac-

tion Planner uses the learning task provided by the

AEG to generate a complete and coherent action plan

for the trainee (see Section 3.2.1). The resulting ac-

tion plan imposes requirements on the game world;

e.g. actions may require the presence of objects or co-

ordination with tasks performed by other characters

(agents). The Object Selector warrants that such re-

quirements are met by selecting the appropriate ob-

jects and agents (see Section 3.2.2).

The Scenario Generator requires the following in-

put from the AEG. 1) The learning task encompass-

ing the learning goal that the trainee is supposed to

AutomatedScenarioGeneration-CouplingPlanningTechniqueswithSmartObjects

Student

Model

Task

Model

World

Knowledge

Base

Adaptive Educational Game

Trainee actions,

Agent goals,

Smart objects

Action

Planner

Object

Selector

Scenario Generator

Learning task,

Diculty level,

Scenario template,

Setting

Learning

goal

selection

Learning task &

diculty level

selection

Performance

evaluation &

feedback

Student

model update

Scenario play

Figure 1: An overview of the Scenario Generator within the Adaptive Educational Game.

practise during the scenario (e.g. ‘treat burn’). 2) The

difﬁculty level at which the learning task should be

practiced. This level is deﬁned as a value between

0 and 1, representing the skill level required to suc-

cessfully perform the learning task, where 0 indicates

no skill and 1 represents mastery. 3) The scenario

template: an ordered list of high-level tasks that con-

stitute a complete training exercise. 4) The setting,

identifying the desired contextual location of the sce-

nario (e.g. ‘the kitchen’). The setting is used to en-

sure authenticity and inﬂuences the object selection

process. Additionally, the Scenario Generator is as-

sumed to have access to a student model containing

the trainee’s current performance levels, a domain-

speciﬁc task model containing information about the

decomposition of learning tasks, and a world knowl-

edge base describing the available objects and agents

along with their domain-speciﬁc features.

The output of the Scenario Generator is a scenario

plan that contains all the information required by the

AEG. 1) Trainee Action Plan: a partially ordered list

of actions the trainee is expected to perform. 2) Agent

Goals: a list of high-level goals for the virtual agents

(e.g. ‘create ﬁre’). 3) Smart Objects: a list of the re-

quired Smart Objects annotated with parameters that

inﬂuence the difﬁculty level at which they offer the

required services (i.e. interaction possibilities).

3.2.1 Action Planner

The Action Planner creates a coherent action se-

quence for the trainee that encompasses the learning

task, and constitutes a complete training exercise be-

cause of the scenario template used in the process. In

addition, the planner determines the goals for the vir-

tual agents (i.e. events) that are expected to trigger

this action sequence. Like Niehaus and Riedl (2009),

the action planner employs a hybrid HTN plan-space

planner. Such a planner employs domain-speciﬁc

knowledge to decompose abstract high-level tasks

into concrete actions meanwhile addressing open pre-

conditions by adding new actions and introducing

causal links between existing actions. The domain-

speciﬁc knowledge used by the HTN planner is stored

in the form of so-called methods. A method speciﬁes

how a high-level task can be decomposed into sub-

tasks (see Ghallab et al. (2004) for a more formal def-

inition). For our particular purpose, we extend the

domain knowledge stored in an HTN method with

two components: preconditions in the form of ser-

vices that must be offered by the game world before

a method can be applied; and an indicator of the dif-

ﬁculty level of the method. An example of a precon-

dition would be that to decompose the task ‘remove

danger’ into ‘extinguish ﬁre’ the service ‘ﬁre’ must

ﬁrst be offered. The difﬁculty level, a value between

0 and 1, allows the Action Planner to inﬂuence the dif-

ﬁculty of the scenario by comparing applicable meth-

ods in terms of complexity. For example, decompos-

ing the task ‘treat arm injury’ into ‘clean wound’ and

‘dress wound’ might be less complex than decompos-

ing the same task into ‘clean wound’, ‘dress wound’

and ‘apply splint’. The difﬁculty level is determined

by domain experts. The following paragraphs will

continue the description of the action planner by de-

tailing the different steps of the scenario generation

process (also depicted in Figure 2).

Incorporate the Learning Task. To ensure a com-

plete training exercise, the scenarios need to contain

the learning task and follow the scenario template.

Therefore, the planner searches for (sequences of)

HTN methods that can be applied to decompose the

high-level task(s) from the scenario template into a

series of subtasks containing the learning task.

CSEDU2013-5thInternationalConferenceonComputerSupportedEducation

Select Task

Select Method

Establish

Preconditions

Apply

Decomposition

Method

Select Object

Object Selector

Action Planner

Learning Task &

Diculty Level

Scenario Plan

Incorporate

Learning Task

Figure 2: Flowchart depicting the planning process of the

Scenario Generator.

Select Task. Once the learning task has been intro-

duced in the planning, the system iteratively addresses

the remaining high-level tasks in the scenario tem-

plate. During this process the system randomly se-

lects the next task to work on in order to promote va-

riety between scenarios with the same input.

Select Method. Each selected task is decomposed

into actions by applying decomposition methods.

This process is guided by two considerations: the pre-

conditions already established by the game world so

far, and the difﬁculty level of the method. First, the

planner selects the methods with the highest number

of established preconditions thereby ensuring the co-

herence of the scenario. If there are multiple candi-

dates, the selection process is guided by the smallest

difference between the method’s difﬁculty level and

the desired difﬁculty level. If the task at hand is the

learning task, the desired difﬁculty level is received

from the AEG, otherwise the desired difﬁculty level

is retrieved from the student model. Any remaining

ties between methods are broken randomly.

Establish Preconditions. Before a method can be ap-

plied, all its preconditions (required services) must

be fulﬁlled. The Action Planner forwards the ser-

vice requests to the Object Selector. Sometimes a

service can only be offered after a certain goal has

been achieved, e.g. a ‘ﬁre’ can only be offered after

it has been ignited. The Object Selector returns these

goals to the Action Planner which plans an action se-

quence for a virtual agent to achieve this goal. NB:

Although the action sequence is not relevant for the

output (the virtual agents require goals only), it must

be computed since the actions may in turn require new

services from the game world.

Apply Decomposition Method. The ﬁnal step in the

process is to actually replace the high-level task with

the subtasks speciﬁed in the selected method. If these

subtasks are actions, the Action Planner also needs to

ensure any open preconditions of the actions.

3.2.2 Object Selector

The Object Selector is grants the service requests

posed by the Action Planner, by reasoning about the

services (i.e. interaction possibilities) the available

Smart Objects can offer, such as a match offering the

service ‘ﬁre’. Agents are considered to be a special

kind of Smart Object that can offer more complicated

services. In addition, Smart Objects are annotated

with preconditions for their offered services, an in-

dication of their belonging in speciﬁc settings, and a

difﬁculty level. The preconditions can be other re-

quired services, or goals that should be achieved by

an agent. The difﬁculty level is based on three as-

pects: the complexity of its use, the obviousness of its

intended use for the service, and the adaptability of its

difﬁcultylevel (i.e. a ﬁre can be small and controllable

or a raging inferno). The Object Selector determines

the most appropriate object based on its belonging in

the provided scenario setting and its match to the de-

sired difﬁculty level. If the selected object requires

any other services, the Object Selector iteratively ful-

ﬁls these requests; if the object requires the achieve-

ment of speciﬁc goals the Object Selector forwards

the request to the Action Planner.

4 EVALUATION

A (proof of concept) implementation of the scenario

generator was evaluated, be it with some additional

simpliﬁcations due to time and resource constraints.

The prototype employs a rudimentary HTN planner

(not an HTN plan-space planner); all service precon-

ditions are deﬁned as actions (not goals), and there are

no ordering constraints on the actions of the agents.

These simpliﬁcations are of no great concern for the

current test setup, though should be addressed in fu-

ture implementations to fully exploit the advantages

of using autonomous agents. The prototype is devel-

oped for the training domain of First Aid, which has

the advantage that it requires no complex world rep-

resentations and has clearly deﬁned procedures. To

further limit the size of the knowledge base the proto-

type was restricted to burn-related incidents.

AutomatedScenarioGeneration-CouplingPlanningTechniqueswithSmartObjects

Table 1: Data exploration: mean scores (and standard deviations) over all raters.

Scenario Source

Dependent Variable

Expert Layman System

Overall

competency suitability

.867 (.448) -.061 (.431) .000 (.527) .269 (.413)

task suitability

1.933 (.290) .389 (.453) .800 (.534) 1.041 (.358)

authenticity

1.733 (.210) .450 (.400) .583 (.199) .922 (.167)

4.1 Method

Evaluators - oblivious of the research question -

rated scenarios written by the system, human ex-

perts and laymen in random order. The sce-

narios were evaluated on three of the require-

ments identiﬁed in Section 3.1: suitability for the

learning task (task suitability), suitability for the

trainee’s competency level (competency suitability),

and authenticity. Completeness of the exercise was

omitted, it being too hard to recognize by just the sce-

nario description (the trainee’s expected action plan

should be included). It was hypothesised that for all

dependent variables the experts would score best, fol-

lowed by the system followed by the laymen. The

setup was a within-subjects design; all evaluators

rated all scenarios from all sources.

Participants. The experiment used 5 evaluators (all

First Aid instructors), and 9 writers (5 First Aid in-

structors and 4 laymen).

Scenarios. The test set consisted of 36 scenarios: 12

scenarios for each source (experts, laymen, and sys-

tem). All authors wrote scenarios based on 3 features:

1) the learning task (i.e. treat burn, calm victim, or

ensure ABC), 2) the trainee’s competency level (i.e.

beginner or advanced), and 3) the setting (i.e. home,

restaurant, laboratory, or park). Counterbalancing

ruled out any possible effects resulting from these fea-

tures. The authors used a predeﬁned format, consist-

ing of the background story (i.e. what happened), in-

structions for the ‘victim agent’, and a list of required

objects. The output of the system was not in natural

language and was manually rewritten following pre-

deﬁned translation rules.

Questionnaires. Each page of the questionnaire

contained a scenario description followed by 3 7-

point Likert-scale questions. The task suitability

and competency suitability were measured indi-

rectly; the evaluators rated the suitability of the sce-

nario for a beginner and for an advanced trainee.

The highest of the two was used to represent the

task suitability. The score of the question matching

the intended difﬁculty level was used as a measure of

the competency suitability.

Procedure. In advance of the experiment proper and

after an extensive instruction, inter-rater reliability

was fostered by a joint discussion on 2 sets of 6 ex-

ample scenarios.

Results. Table 1 shows the means and standard

deviations of the test set. The results follow the

hypothesised trend with the experts scoring high-

est followed by the system followed by the lay-

men. The table also shows large standard devia-

tions, in particular for the system and layman sce-

narios. The intraclass correlation coefﬁcient (ICC)

using the 2-way random model suggested substan-

tial agreement (r=0.732; p<.001). Any missing val-

ues (0.01% of the 540 values) were imputed using

the SPSS expectation-maximisation procedures (Lit-

tle and Rubin, 1989). A repeated measures ANOVA

revealed signiﬁcant differences between the sources

for task suitability (F(2, 8) = 6.699; p = .020) and

authenticity (F(2, 8) = 6.220; p = .023), but not for

competency suitability (F(2, 8) = 3.529; p = .080).

Post-hoc tests using the Bonferroni correction proce-

dure revealed no signiﬁcant differences between the

sources in one on one comparisons.

Discussion. The experiment revealed no signiﬁcant

differences between the scenarios coming from dif-

ferent sources (i.e. experts, laymen, and system). One

possible explanation might have been the large stan-

dard deviations, which could havebeen caused by dis-

agreement among the evaluators, however, the ICC

analysis rules out that possibility. Several points for

discussion that may shed some light on these results

are discussed. First of all, the settings were purpose-

fully varied to show that the prototype is capable of

generating scenarios for various situations. However,

these settings may have forced the experts to write

scenarios for settings they normally would not use.

Secondly, the current template does not include the

unfolding of the scenario, nor the trainee’s expected

action sequence, thereby failing to reveal the source’s

intentions underlying the produced scenarios. The

evaluators may have used their experience to interpret

the scenario descriptions to infer these aspects, cover-

ing up any possible differences. A ﬁnal point of atten-

tion is the importance of involving domain experts in

the development of the knowledge base: the task de-

compositions and the objects used. Even though the

used knowledge base contained accurate information,

the instructors indicated that it contained some con-

tent they would never use in their own scenarios, e.g.

electrical burns.

CSEDU2013-5thInternationalConferenceonComputerSupportedEducation

5 DISCUSSION & CONCLUSIONS

This paper addressed the issue of automated scenario

generation within the context of an Adaptive Educa-

tional Game (AEG). The proposed framework inte-

grates a hybrid HTN planner to plan the trainee’s ac-

tions with a content selection mechanism based on

Smart Objects to control the realisation of the sce-

nario within the game world. This results in a separa-

tion of the action plan construction - enabling the re-

trieval of the scenario’s underlying didactic intentions

- and the game world creation - enabling the use of

a separate smart objects database which is easily ex-

tended, in contrast to hard-coded objects in the plan-

ner itself. Although the evaluation experiment did

not provide signiﬁcant results, the authors are hopeful

that further experimentation will provide more deﬁni-

tive answers.

Several directions for further research can be sug-

gested. First of all, the experiment showed that some

of the desired functionalities of the system interfered

with its primary goal of producing scenarios with a

complexity level that ﬁts the skill level of the trainee.

Since the prototype generated complete training exer-

cises, additional tasks were addressed in the scenario

on top of the learning task. Because a global difﬁ-

culty level was used, the trainee is expected to per-

form the additional tasks at the same difﬁculty level as

the learning task. This combination of multiple tasks

seemed too much to handle for the trainee. Two pos-

sible (and complimentary) solutions are the integra-

tion of a more ﬁne-grained difﬁculty control system

such as for example a performance curve suggested

by Zook et al. (2012) and the introduction of ‘col-

league agents’ that take over part of the responsibili-

ties of the trainee. A second suggestion for further re-

search is the comparison of the different approaches

proposed for automated scenario generation. So far

it has been difﬁcult to compare different approaches

since each system uses its own standards and criteria.

However, a comparison might be highly informative

and show the strengths and weaknesses of the differ-

ent approaches, possibly leading to hybrid solutions.

To conclude, the framework presented here is a

ﬁrst step towards a system that can generate effective

and personalized scenarios automatically. In the fu-

ture, attaching the system to a game-based training

program will offer trainees extensive access to high-

quality training opportunities.

REFERENCES

Ghallab, M., Nau, D., and Traverso, P. (2004). Automated

Planning: Theory & Practice. Morgan Kaufmann

Publishers Inc., San Francisco, CA, USA.

Grois, E., Hsu, W. H., Voloshin, M., and Wilkins, D. C.

(1998). Bayesian network models for generation of

crisis management training scenarios. In Proc. IAAI-

98, volume 10, pages 1113–1120. AAAI.

Issenberg, S., McGaghie, W., Petrusa, E., Gordon, D., and

Scalese, R. (2005). Features and uses of high-ﬁdelity

medical simulations that lead to effective learning: a

beme systematic review. Medical Teacher, 27(1):10–

28.

Kallmann, M. and Thalmann, D. (1998). Modeling objects

for interaction tasks. In Proc. Eurographics Workshop

on Animation and Simulation, volume 9, pages 73–86.

Citeseer.

Little, R. and Rubin, D. (1989). The analysis of social sci-

ence data with missing values. Sociological Methods

& Research, 18(2-3):292–326.

Lopes, R. and Bidarra, R. (2011). Adaptivity challenges

in games and simulations: A survey. IEEE Transac-

tions on Computational Intelligence and AI in Games,

3(2):85–99.

Martin, G., Schatz, S., Bowers, C., Hughes, C. E., Fowlkes,

J., and Nicholson, D. (2009). Automatic scenario

generation through procedural modeling for scenario-

based training. In Proc. HFES-09, volume 53, pages

1949–1953. SAGE Publications.

Merrill, M. D. (2002). First principles of instruction.

ETR&D, pages 43–59.

Niehaus, J. and Riedl, M. O. (2009). Scenario adaptation:

An approach to customizing computer-based training

games and simulations. In Proc. Workshop on Intelli-

gent Educational Games (AIED-09), volume 3, pages

89–98.

Oser, R. L., Cannon-Bowers, J. A., Salas, E., and

Dwyer, D. J. (1999). Enhancing Human Perfor-

mance in Technology-Rich Environments: Guide-

lines for Scenario-Based Training, volume 9 of

Human/Technology Interaction in Complex Systems,

pages 175–202. Jai Press Inc.

Peeters, M., Bosch, K., Meyer, J. J., and Neerincx, M.

(2012). Situated cognitive engineering: the require-

ments and design of directed scenario-based training.

In Proc. ACHI-12, volume 5, pages 266–272. XPS.

Peirce, N., Conlan, O., and Wade, V. (2008). Adaptive

educational games: Providing non-invasive person-

alised learning experiences. In Proc. DIGITEL-08,

volume 1, pages 28–35. IEEE.

Salas, E., Priest, H. A., Wilson, K. A., and Shawn Burke,

C. (2006). Scenario-Based Training: Improving Mili-

tary Mission Performance and Adaptability, volume 2

of Military Life: the psychology of serving in peace

and combat, chapter 3, pages 32–53. Praeger Security

International.

Zook, A., Urban, S. L., Riedl, M. O., Holden, H. K., Sotti-

lare, R. A., and Brawner, K. W. (2012). Automated

scenario generation: toward tailored and optimized

military training in virtual environments. In Proc.

FDG-12, volume 7, pages 164–171. ACM.

AutomatedScenarioGeneration-CouplingPlanningTechniqueswithSmartObjects