Thinking With Containers: A Multi-Agent Retrieval Approach for the

Case-Based Semantic Search of Architectural Designs

Viktor Ayzenshtadt

1,2

, Christoph Langenhan

, Saqib Bukhari

Klaus-Dieter Althoff

1,2

, Frank Petzold

and Andreas Dengel

University of Hildesheim, Institute of Computer Science, Samelsonplatz 1, 31141 Hildesheim, Germany

German Research Center for Artiﬁcial Intelligence (DFKI), Trippstadter Strae 122, 67663 Kaiserslautern, Germany

Faculty of Architecture, Technical University of Munich, Arcisstrasse 21, 80333 Munich, Germany

Keywords:

Multi-Agent Systems, Case-Based Agents, Case-Based Retrieval, Design Support.

Abstract:

To provide the retrieval of information, that can be considered useful during the design conceptualization

process, with advantages of distributed artiﬁcial knowledge, an approach, that distributes retrieval-related and

knowledge maintaining tasks among autonomously working and case-based self-learning agents and agent

groups, can be used. In this work we present the distributed retrieval system MetisCBR for the architectural

design domain, where agents work in groups (containers) on resolving of user queries built with a semantic

description model Semantic Fingerprint. The main aim of our approach is to carry out a basis for a considerable

retrieval tool for architects, where the combination of case-based reasoning and multi-agent methods helps to

achieve valuable and helpful search results in a comprehensive building design collection.

1 INTRODUCTION

The early conceptualization phase of a building is the

stage of the architectural design development, during

which an architect can increase the efﬁciency of the

working process by communicating with a computer-

aided helper system. Such a system should be able to

present new ideas and helpful recommendations for

inspiration and guidance in combination with similar

building designs to a currently composed one.

The study of buildings that have a similar con-

text to the design problem at hand, or that are based

on a similar initial premise, is seen as a way of ap-

proaching a design problem and developing a possible

course of action. Prior built projects or architectural

designs serve here as a knowledge base that contains

both examples of spatial constellations as well as so-

lutions for speciﬁc architectural situations.

Plans, models, and the documentation of architec-

tural designs and existing buildings and urban envi-

ronments represent an extensive architectural knowl-

edge base which contains implicit design knowledge

that is valuable for the design process. The problem

is, that this knowledge is not available in an explicit

form, without ﬁrst analyzing and interpreting the in-

formation available.

To handle this architectural knowledge, for the

purpose of providing a retrieval component that ef-

ﬁciently interprets this data, the helper systems of-

ten use case-based reasoning (CBR) methods as the

underlying retrieval technique. This aims to achieve

high concentration on the quality of results, and to

make the system knowledge-intensive and extensible.

For implementation of such retrieval features,

many approaches were proposed and applied to the

existing systems in the last three decades. The

most signiﬁcant of them are listed in (Heylighen and

Neuckermans, 2001) and (Richter et al., 2007).

In this work we present MetisCBR, a multi-agent

and case-based retrieval approach for similar archi-

tectural designs. This approach uses inter alia case-

based learning agents as retrieval units in a distributed

container-based application. MetisCBR was imple-

mented as a prototype and integrated into the existing

infrastructure in context of a basic research project

Metis – Knowledge-based search and query methods

for the development of semantic information models

(BIM) for use in early design phases, an interdisci-

plinary project of CBR, multi-agent systems (MAS),

and computer-aided architectural design (CAAD),

partially funded by the German Research Foundation

(Deutsche Forschungsgemeinschaft, DFG).

This paper is structured as follows: ﬁrst, the back-

ground and the interdisciplinarity of the research

Ayzenshtadt, V., Langenhan, C., Bukhari, S., Althoff, K-D., Petzold, F. and Dengel, A.

Thinking With Containers: A Multi-Agent Retrieval Approach for the Case-Based Semantic Search of Architectural Designs.

DOI: 10.5220/0005683801490156

In Proceedings of the 8th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2016) - Volume 1, pages 149-156

ISBN: 978-989-758-172-4

149

project Metis will be brieﬂy presented. The next sec-

tion contains the overview of related approaches and

studies. Next, a short overview of the system struc-

ture will be presented. The next section character-

izes the proposed agent types and their correspond-

ing tasks in detail. In the next section the prototype

of the approach and its context and validation within

the research project will be presented. The conclusion

summarizes the work and the future project planning.

2 BACKGROUND

Technological developments and globalized working

processes have transformed the building design pro-

cess. However, current digital semantic building

models are no longer able to adequately represent

the increasing complexity of modern architectural

projects. The potential of combining agent-based,

case-based, and graph-based methods with novel se-

mantic models, for example, for energy calculations

or spatial research strategies, is not fully exploited.

To ﬁnd essential ways to utilize this potential, a basic

research project Metis was initiated. The distributed

CBR retrieval system MetisCBR we present in this

work is one of the integrant parts of the project. The

interdisciplinary nature of the project provides a nec-

essary mixture to overcome the problem of timely

ﬁnding similar building designs as source of inspira-

tion to guide the design process of buildings.

CAAD considers computer-related topics regard-

ing buildings and delivers insights into ways an archi-

tect can work/interact with such a system or how she

would evaluate search results.

CBR, as the discipline responsible for working

with intensive knowledge data, is the means for the

actual ﬁnding of similar building designs. CBR’s all-

purpose nature and a multitude of different CBR re-

trieval approaches provide efﬁcient tools for the ac-

complishment of the project’s main tasks.

MAS play a role of executing units in the project.

Being a means for distributed decentralized problem

solving, MAS, and the corresponding research area,

provide a necessary collection of techniques for dis-

tributed and concurrent retrieval processes.

To combine the above named ﬁelds in order to

work together in a common system, a building in-

formation model (BIM) infrastructure was built for

the project. It integrates every project-related service,

including a building model server, a graph database

with building designs represented as graphs, a content

management system, and software prototypes.

3 RELATED WORK

The system we present in this work is set in context

of the above mentioned research project Metis, and is

related to the CAAD and BIM applications with case-

based reasoning as the underlying retrieval technique.

The project-related work includes (Siebert, 2014),

a research work that contains the initial setting of our

system where a concept of agent-based experience

sharing system is applied to the architectural domain.

A semantic building structure description model

Semantic Fingerpint (Langenhan and Petzold, 2010)

was initiated to provide a structure to idea that build-

ing ﬂoor plans can be described by abstracting the

building data at the level of building, storey, and

room, as well as their respective topological relation-

ships. Based on this model, a subgraph matching-

based retrieval algorithm was proposed in (Ahmed

et al., 2014), where an information extraction process

of ﬂoor plans is also contained and later used in the

Metis project for the graph database.

Applications for Android and iPad, and a touch-

table interface, were developed for query construction

with modern haptic HCI methods. A web-based user

interface (Bayer et al., 2015) is part of the project as

well. This application uses the (AGraphML) speciﬁ-

cation (Langenhan, 2015) for the query construction.

Finally, a distributed domain model that builds a

CBR base for the retrieval process of MetisCBR was

introduced in (Ayzenshtadt et al., 2015).

CBR-based approaches with architectural empha-

sis can be classiﬁed as a subdomain of case-based de-

sign (CBD). In these systems, the retrieval function-

ality is one of the most essential system modules.

For example, in FABEL (Voss, 1997) functional

entities called specialists are used to execute the

retrieval tasks. The specialists work with special

aspect-speciﬁc representations of cases and queries.

These representations are produced by applying a

transformation process to the user query or to a cur-

rently comparing case of the case base. Retrieval spe-

cialists determine similar cases by comparing the cor-

responding aspects of cases and queries.

CBArch (Cavieres et al., 2011) is a supporting

framework for the conceptualization phase of archi-

tectural design. With emphasis on the early energy

performance evaluation of buildings with commercial

background, it provides an approach for a completely

CBR-based solution for such cases.

Other systems, such as DYNAMO (Heylighen

et al., 2002), SEED (Flemming, 1994), PRECE-

DENTS (Oxman and Oxman, 1993) or CaseBook

(Inanc, 2000), can also be mentioned as closely re-

lated, as they serve the same purpose as our system.

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

150

Web UI

CBR domain model

(myCBR API)

Graph

Semantic

ngerprints

(extracted from sketches

with image processing)

GraphML IO

Semantic

ngerprint

Case

Bases

Coordinator

CBR agent 1

CBR agent 2

CBR agent 3

More CBR agents ...

Room µ

Edge µ

Maintainer agent

GraphDB agent

User query

GraphML agent

Ontological

query

Cases to import

Semantic FP

Ontological

query

Search result

Sets the retrieval strategy based on

case-based and/or rule-based reasoning

and starts a new retrieval container

Floorplan µ

Checks the DB for new

graphs and imports them

into the case base

Sends

Semantic FP

Looks up for the similar

graphs in the DB by

querying it and sends

the proper data back

Consructs an ontological query

from the XML/GraphML of the

Semantic ,ngerprint

Gateway

Processes

Semantic FP

CBR manager

Ontological

query

Retrieval strategy

More retrieval containers ...

Retrieval container

Separates the query into

the sequences and

collects result data

Graph DB query

Figure 1: The system overview of MetisCBR.

4 THE SYSTEM OVERVIEW

In this section we provide a brief overview of the pro-

posed approach MetisCBR. The main task of the en-

tire multi-agent system is to process a query in the

AGraphML format that was sent by a user (an archi-

tect) from a web-based query construction interface

(not part of this approach, see also Sections 3 and 6).

As depicted in Figure 1, the system is constructed

in a circular form, where the gateway marks the entry

point into the system, after a user query was sent from

the web interface.

Successfully passing the gateway agent that con-

ducts an initial error analysis, the query reaches the

coordination component that is responsible for the

correct accomplishment of the entire case-based re-

trieval process. Its ﬁrst action is to send the query to

the parsing agent that transforms it from the architec-

tural GraphML into an ontological representation in

SL content language to provide human-readability as

well as suitability for communication among agents

according to (Caire and Cabanillas, 2002).

Back at the coordination point, the query is analyzed

and sent to a retrieval container that conducts the ac-

tual resolving of this particular query. The coordina-

tor selects the proper retrieval strategy and conducts

the setup of the container, where a retrieval manager

and at least one case-based retrieval agent that oper-

ates with at least one similarity measure µ are manda-

tory units for successful processing of the retrieval

stage in the case base(s) of the CBR domain model.

After ﬁnishing the actual retrieval process, the

achieved result data is sent back from the container’s

manager to the coordinator, who forwards it to the

gateway that transforms it into the proper processing

format (JSON) and sends the result data object to the

web interface component for displaying.

Parallel to the user query resolving process, new

cases (building designs) are being imported into the

case base(s) of the underlying CBR domain model.

This is done by another essential component of the

proposed MAS, the maintainer, whose main task is

to extract cases from the BIM infrastructure’s graph

database and to import them into these case base(s).

Thinking With Containers: A Multi-Agent Retrieval Approach for the Case-Based Semantic Search of Architectural Designs

151

5 AGENT TYPES AND TASKS

This section presents detailed descriptions of the

agent types that are directly or indirectly involved in

the process of retrieval of similar building designs.

Their tasks will be also described in detail.

5.1 Agent Categories

The categorization of agents into classiﬁed groups

helps to achieve a better understanding of the entire

system functionality. For our case-based multi-agent

approach with retrieval as main task, we deﬁned the

following categories of agents: case-based, manag-

ing, connecting and service agents. An agent may be-

long to one or many categories.

Case-based agents are directly involved in the re-

trieval process of similar architectural designs and

possess an internal reasoning and learning component

with rule-based and/or case-based reasoning abilities.

Retrieval executing agents of this category (see Sec-

tion 5.3.2) are able to apply different similarity mea-

sures when searching for designs (or parts of it) in the

case base(s) of the CBR domain model described in

(Ayzenshtadt et al., 2015) (see also Sections 3 and 6).

Managing agents ensure that the entire system (in-

cluding the retrieval process and building designs im-

port into the case base(s)) is working properly, with-

out failures, such as abruptions or sudden bottlenecks.

The agents of this class monitor ongoing internal pro-

cesses, share their statuses/intentions, and forward

queries and results to the corresponding agents.

Connecting agents play the role of connection

points between the MAS and linked external services.

They make sure that the connection and communica-

tion of MetisCBR with these environments (e.g., the

query builder interface or the graph database) is es-

tablished, and query or case streams can be forwarded

to the destinated system components.

Service agents are helper agents that execute tasks

that do not belong directly to retrieval, management,

or connection task domains. The agents of this class

are involved in the accomplishment of intermediate

steps of the retrieval process (e.g., query parsing).

5.2 Coordination

For the efﬁcient coordination of the retrieval process

within our system, a corresponding agent (the coordi-

nator) was implemented. Its role is to manage the en-

tire case-based retrieval stage, after the user query has

been received from the gateway. Therefore, this agent

belongs to both categories of case-based and manag-

ing agents, and has a number of particular tasks.

Retrieval Strategy Selection. The selection of the

proper retrieval strategy is the coordinator’s most es-

sential task. It is based on the combination of the

properties of the query, previous experiences, rules,

and user-speciﬁed data from the query.

For example, if the user’s current objective is to

conduct a retrieval without much detailed settings for

the search, the coordinator can derive the objective

from the user-speciﬁed data (in form of a speciﬁc at-

tribute) and select the associated retrieval strategy to

resolve the query. Otherwise, the common strategy

can be applied. The objective recognition and setting

is an example of applying the rule-based reasoning,

the ﬁrst one of the reasoning abilities of the coordi-

nator. The second one is the case-based classiﬁcation

of the query, where selection of the strategy is mainly

the task of the learning component of the coordinator.

Learning is a signiﬁcant property of an agent that

aims to improve its ability to make decisions based

on previous experiences. Therefore, the coordination

component of our retrieval system, being such an

agent, uses its case-based learning ability as an

essential feature to select the proper procedure for

query resolving. Each query that the coordinator gets

from a user, together with the achieved results, is

registered as a case with the IB2 algorithm (Aha et al.,

1991) that controls the ﬁlling of the coordinator’s

internal case base and provides the coordinator with

the learning ability. These cases are the base for the

reasoning process that gets activated if the rule-based

classiﬁcation misses the explicit speciﬁcation from

the user side. During the CBR-based classiﬁcation,

the corresponding ﬂoor plan information of the query

(e.g., room count and room types) is used to ﬁnd the

most similar case in the coordinator’s internal case

base, and apply a strategy that was used in this case.

Retrieval Container Activation. After the strategy

for the retrieval has been selected, the coordinator ac-

tivates a new retrieval container (see Section 5.3). The

setting of the container is deﬁned by the coordinator

– it selects the retrieval agents that will execute the

actual retrieval steps and starts the activity of the con-

tainer. In MetisCBR the retrieval strategy is essen-

tial for selection of the agents – retrieval agent types

are bound to the strategy, but not relevant for the se-

lection of the strategy. The coordinator is also able

to terminate one or more currently running retrieval

containers. This happens when a container has ﬁn-

ished its retrieval activities, and the user has ﬁnished

her retrieval session. In case of unexpected interrup-

tion of the connection between the user interface and

the MAS, the retrieval is continued and the results are

persisted in an intermediate object.

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

152

Forwarding of the Query Modiﬁcations. If it is

the user’s wish to modify the current query during

the ongoing retrieval, the coordinator receives the

modiﬁed query from the gateway agent and forwards

the changes while notifying the manager of the

corresponding container about these changes.

Sending of the Results to the Gateway. When the

actual retrieval phase is ﬁnished, the coordinator

receives the results from the manager of the container

and provides the gateway agent with these results.

Assignment of the Query Parsing Task. The

coordinator assigns the GraphML agent (see Section

5.6) the task of parsing the current query, which it

gets back transformed as an ontological query in SL

content language format.

Receiving the Import Status. From the maintainer

agent the coordinator receives the current status of

the extraction/import process and coordinates the cur-

rently running retrieval tasks with respect to this sta-

tus (see also Section 5.5).

5.2.1 Coordination Techniques

The efﬁcient accomplishment of the above described

tasks requires possession of corresponding coordina-

tion techniques. The coordination component of our

retrieval system is developed with the integration of

combination of some of the coordination scenarios

described in (Nwana et al., 1996).

Organizational Structuring is applied when the

coordinator creates and destroys retrieval containers

and assigns strategies to them. It is the main coordi-

nation technique in MetisCBR.

Contracting (more precisely, a sub-feature of con-

tracting, the sub-distribution of the tasks) is applied

when a manager of a retrieval container assigns parts

of the query to be resolved by corresponding agents.

Centralised Multi-agent Planning is activated when

the maintainer sends an import status to the coordina-

tor (see Section 5.5)

5.3 Retrieval Container

In a retrieval system, where concurrent queries are ex-

pected to be a usual case, a big advantage is to im-

plement a structure that is able to handle each of the

queries separately to achieve more efﬁcient process-

ing. For the actual building design retrieval step of our

approach, we use a container-based structure, where

each retrieval container during its activity can be seen

as a completely autonomous multi-agent sub-system

that is nested inside a parent MAS, and can only be

terminated if its retrieval process and the user session

are both ﬁnished.

The working cycle of a retrieval container starts

with receiving of an ontological query from the co-

ordinator and ends after the retrieval of the assigned

query is ﬁnished. Between those two steps, a case-

based resolving of the query takes place inside the

container. Thus, every container has three goals to

complete during its working cycle: receiving of the

query, resolving of the query by means of applying

CBR methods, and replying to the coordinator with a

message that contains the achieved results.

Case

base

Coordinator

CBR Agent 1

CBR Agent 2

CBR Agent 3

Room µ

Edge µ

Floorplan µ

CBR Manager

→ Ontological query

← Retrieval results

Retrieval container

Figure 2: Exemplary conﬁguration of a retrieval container.

An exemplary conﬁguration of a single retrieval con-

tainer can be seen in Figure 2. Every container con-

sists of a mandatory CBR manager – an agent that

manages resolving of the current query – and at least

one mandatory CBR retrieval agent – an entity that

is responsible for the actual search for similar cases.

The following sections provide a detailed description

of both agent types.

5.3.1 CBR Manager

As mentioned above, the CBR manager is a manda-

tory agent inside an active retrieval container. It orga-

nizes the current case-based retrieval of an ontological

query that was received from the coordinator. There-

fore, this agent belongs to the category of managing

agents (see Section 5.1), and was primarily created to

take load off the coordinator, that can in turn concen-

trate on its strategic and system managing tasks.

The tasks of the manager include two primary

classes: communication with the coordinator and

query separation. During the communication with the

coordinator, (i.e., receiving an ontological query and

sending the results back), the manager completes re-

ceiving and replying steps of the container goals. The

separation process cuts the given query into the num-

ber of parts according to the selected retrieval strategy

and forwards these parts to the corresponding agents.

Thinking With Containers: A Multi-Agent Retrieval Approach for the Case-Based Semantic Search of Architectural Designs

153

The secondary class of tasks of the manager consists

of helper operations, which are accomplished only if

required by the strategy as well. Those helper op-

erations can include some computational tasks, such

as determination of rooms and room connections that

originate from the same ﬂoor plan.

5.3.2 CBR Retrieval Agents

CBR retrieval agents execute the last and most

important part of the retrieval process – they apply

CBR retrieval methods for determination of the most

similar cases to the query or its parts in the case

base(s) of the underlying attribute-value based CBR

domain model (described in detail in (Ayzenshtadt

et al., 2015)). To achieve this goal, they use sim-

ilarity functions prepared for each agent, special

query types, and an internal reasoning mechanism

to deal with incoming queries from the manager.

CBR retrieval agents belong to the case-based agents,

where several agent types of this class currently exist.

Floor plan Agent. This agent’s aim is to search

for cases that contain ﬂoor plan meta data from

the AGraphML speciﬁcation and additional related

information such as room count and room types.

Floor plan agent works with following query types:

• Floorplan Query – every attribute of the meta

data and additional data will be taken into account

for the similarity measurement.

• Floorplan ID List Query – returns cases, if

their corresponding IDs exist in a deﬁned ID list.

Room Agent. To ﬁnd similar rooms to a given one

this agent uses every room attribute from the domain

model. Its only query type is currently Room Query

that accomplishes this similarity computation task.

Edge Agent. This retrieval agent takes every edge

attribute into account to compute a similarity measure

between two room connections. The corresponding

query type is the Edge Query.

5.3.3 Learning of CBR Retrieval Agents

CBR retrieval agents are able to learn from previous

queries in the same manner as the coordinator does.

Each of the retrieval agents possesses an internal rea-

soning mechanism, where a special internal CBR do-

main model exists for each retrieval agent’s knowl-

edge base. Special attributes are determined to hold

information about previous queries the agent has re-

solved (e.g., room count from the query and from the

found case(s) for the Room agent).

5.3.4 Container Identiﬁcation

Retrieval container can be identiﬁed via an UUID as

deﬁned in the RFC 4122. The coordinator assigns an

UUID to each container it creates, the same ID is also

temporarily assigned to the query that the container is

working with to be able to forward modiﬁcations to

the associated container.

5.4 Gateway

As described above in the overview section, the gate-

way agent is an entry point to the MAS from the user

side, and an interface for forwarding queries and re-

sults in corresponding directions. The gateway agent

belongs to the category of connecting agents (see Sec-

tion 5.1).

Gateway agents of such type are common in sys-

tems with structure similar to ours. For example, in

(Greenwood et al., 2005) a gateway agent is a part of a

component that is used for the OWL-S-based seman-

tic interaction between a web client from an external

web service environment and an agent environment.

The activity area of our gateway agent is divided

into two main sections: accepting of user queries and

forwarding them to the coordinator, and receiving re-

sults from the coordinator and forwarding them to the

web user interface. Both task sections are controlled

by special agent behaviors (see Figure 3).

Receiver

behavior

Sender

behavior

Case-based

retrieval

Web UI

Query

Result

Figure 3: The behaviors structure of the gateway agent.

Sending a formalized and validated user query in

AGraphML format to the coordinator is the main pur-

pose of creating of the Sender behavior. During the

validation of the query the gateway agent traverses

through the data in order to ﬁnd formal errors, such

as missing meta data or false (or missing) room or

room connection deﬁnitions. Empty queries will not

be forwarded to the coordinator. Incomplete queries,

where data is fragmentary or only partly missing, will

be forwarded to the coordinator, however, missing at-

tributes will not be used for the retrieval.

Receiving result data from the coordinator and deliv-

ering it to the web-based user interface is the activity

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

154

ﬁeld of the Receiver behavior of the gateway agent.

Similar to the process of the sender behavior, the re-

sult data (in SL content language) will be validated as

well, following by the transformation into the JSON

format (plain HTML is also possible). It is optional if

results will be returned in sequences or at once.

5.5 Maintainer

Maintainer is an agent that is responsible for the cor-

rect functionality of the extraction and import of the

graph-based designs from the BIM infrastructure’s

graph database into the case base(s) of the underlying

case-based domain model of the system. The main-

tainer belongs to the class of managing agents.

During the three-step case base(s) update process

the maintainer executes tasks deﬁned for each step.

In the ﬁrst phase it queries the graph database for

new graph representations of building designs. In the

second phase, the maintainer requests the RESTful

service (implemented in the BIM infrastructure) for

each graph of the previous result set and gets back

an AGraphML-formatted graph. In the last step, the

building graph is imported to the case base(s) of the

domain model with a special import function.

For the essential workﬂow of the system, the

maintainer is also able to set and share an import

status, which currently can be one of the following:

Import Ready, Import Active, Import Paused,

and Import Done.

The coordinator receives the status and may exe-

cute actions according to it. For example, if the sys-

tem performance is critical the coordinator may stop

some retrieval containers or request to stop recently

started extraction/import process to allow for a better

system resource sharing.

5.6 GraphML Agent

As previously mentioned (see Section 5.2), the

GraphML agent is the system entity that is re-

sponsible for parsing of the incoming architectural

GraphML queries into ontological queries in SL con-

tent language format. The ontological queries get

their name from the underlying communication on-

tology that deﬁnes the communication vocabulary

among agents. By means of applying an ontology

to the communication, the ontological query becomes

part of the message content as demonstrated in (Caire

and Cabanillas, 2002).

6 SYSTEM EVALUATION

For the above mentioned basic research project Metis,

a working prototype of the proposed retrieval system

MetisCBR was developed to demonstrate its suitabil-

ity for the ”real-world” application. Java language

based frameworks JADE (for the implementation of

the described agent system) and myCBR were used to

develop the prototype. It is also implemented as one

of the sub-systems of the BIM infrastructure.

The prototype works with the BIM infrastructure’s

content management system (that contains graphical

representations of building designs) and from it de-

rived graph database that contains every building de-

sign as graph representation that is then extracted and

imported as speciﬁed in Section 5.5. In Section 3

mentioned web-based user interface is planned to be

used as the main graphical architectural GraphML

query builder for the retrieval system. Furthermore,

the proposed system and the software prototype are

part of the master thesis (Ayzenshtadt, 2015).

For the validation and demonstration of the abili-

ties of the approach, an evaluation of the system was

conducted, that is completely described in (Ayzensh-

tadt et al., 2015). The evaluation process included

the querying of the system with an exemplary query

consisting of 5 rooms and 6 connecting edges with

two different similarity measures: with and without

attribute weighting.

The MAS task area of the evaluation should show

that MetisCBR is already able to handle search re-

quests and return results according to the similarity

measures.

Considering the early phase of the system imple-

mentation, the agents were able to properly handle

the user queries. The coordinator was able to recog-

nize a new query, to create a corresponding container,

and to assign the strategy and the query type. The

manager of the container was able to recognize if it

is required to separate the query, and then assigned

the corresponding parts to the CBR retrieval agent(s).

The retrieval agent(s) retrieved the case base with the

similarity measure that was set for the current query

(manually, on this stage of development). The man-

ager collected the results and sent them to the coordi-

nator that then forwarded them to the gateway agent.

A couple of minor technical problems occurred dur-

ing the import process, but could be traced by inspect-

ing the consistency service of the BIM infrastructure.

The results of the evaluation of the CBR task area,

including the information about retrieval strategies,

similarity measures, building design dataset used, av-

erage and highest similarity, and the similarity distri-

bution, are described in (Ayzenshtadt et al., 2015).

Thinking With Containers: A Multi-Agent Retrieval Approach for the Case-Based Semantic Search of Architectural Designs

155

7 CONCLUSION

In this paper we presented MetisCBR, a distributed

case-based approach for the retrieval of similar build-

ing designs in the corresponding case base(s). The

system consists of different agent types, where the

coordinator controls the retrieval process and ap-

plies rule-based or case-based reasoning to select the

proper retrieval strategy, retrieval containers execute

concurrently the actual case-based retrieval processes,

the gateway agent connects the system with an exter-

nal service (web-based query builder) and the main-

tainer agent manages the import of new cases into the

case base(s) of the underlying CBR domain model.

Service agents contribute to the smooth run of the sys-

tem by executing small intermediate tasks.

The future work includes the full integration of the

graphical user interface to the presented MAS. Index-

based retrieval method that queries the graph database

directly is also planned to be added, a new GraphDB

agent will be integrated to the MAS to accomplish

this alternative retrieval task. This will also allow for

evaluating and hopefully deeply integrating these dif-

ferent kinds of retrieval approaches.

ACKNOWLEDGEMENT

This work for the Metis project was supported by the

DFG (German Research Foundation).

REFERENCES

Aha, D. W., Kibler, D., and Albert, M. K. (1991).

Instance-based learning algorithms. Machine learn-

ing, 6(1):37–66.

Ahmed, S., Weber, M., Liwicki, M., Langenhan, C., Den-

gel, A., and Petzold, F. (2014). Automatic analysis

and sketch-based retrieval of architectural ﬂoor plans.

Pattern Recognition Letters, 35:91–100.

Ayzenshtadt, V. (2015). Concept and Implementation

of a Distributed Case-based Retrieval Approach for

the Support of Architectural Conceptualization Phase.

Master thesis. University of Hildesheim.

Ayzenshtadt, V., Langenhan, C., Bukhari, S. S., Althoff,

K.-D., Petzold, F., and Dengel, A. (2015). Dis-

tributed domain model for the case-based retrieval

of architectural building designs. In Petridis, M.,

Roth-Berghofer, T., and Wiratunga, N., editors, Pro-

ceedings of the 20th UK Workshop on Case-Based

Reasoning. UK Workshop on Case-Based Reason-

ing (UKCBR-2015), located at SGAI International

Conference on Artiﬁcial Intelligence, December 15-

17, Cambridge, United Kingdom. School of Com-

puting, Engineering and Mathematics, University of

Brighton, UK.

Bayer, J., Bukhari, S., Dengel, A., Langenhan, C., Al-

thoff, K.-D., Petzold, F., and Eichenberger-Liwicki,

M. (2015). Migrating the classical pen-and-paper

based conceptual sketching of architecture plans to-

wards computer tools - prototype design and evalua-

tion. 11th IAPR International Workshop on Graphics

Recognition - GREC’15, Nancy, France.

Caire, G. and Cabanillas, D. (2002). Jade tutorial:

application-deﬁned content languages and ontologies.

TILab SpA.

Cavieres, A., Bhatia, U., Joshi, P., Zhao, F., and Ram, A.

(2011). CBArch: A case-based reasoning framework

for conceptual design of commercial buildings. Ar-

tiﬁcial Intelligence and Sustainable Design – Papers

from the AAAI 2011 Spring Symposium (SS-11-02),

pages 19–25.

Flemming, U. (1994). Case-based design in the SEED sys-

tem. Automation in Construction, 3(2):123–133.

Greenwood, D., Nagy, J., and Calisti, M. (2005). Seman-

tic enhancement of a web service integration gateway.

In Proceedings of the Workshop on Service-Oriented

Computing and Agent-Based Engineering (SOCABE

2005), Utrecht, The Netherlands.

Heylighen, A. and Neuckermans, H. (2001). A case base of

case-based design tools for architecture. Computer-

Aided Design, 33(14):1111–1122.

Heylighen, A., Schreurs, J., and Neuckermans, H. (2002).

Enter instead of submit. DesignNet-Knowledge e In-

formation Management per il design, pages 171–182.

Inanc, B. S. (2000). Casebook. An information retrieval

system for housing ﬂoor plans. In the Proceedings

of 5th Conference on Computer Aided Architectural

Design Research (CAADRIA), pages 389–398.

Langenhan, C. (2015). A federated information system for

the support of topological bim-based approaches. Fo-

rum Bauinformatik Aachen.

Langenhan, C. and Petzold, F. (2010). The ﬁngerprint

of architecture-sketch-based design methods for re-

searching building layouts through the semantic ﬁn-

gerprinting of ﬂoor plans. International electronic

scientiﬁc-educational journal: Architecture and Mod-

ern Information Technologies, 4:13.

Nwana, H. S., Lee, L. C., and Jennings, N. R. (1996). Co-

ordination in software agent systems. British Telecom

Technical Journal, 14(4):79–88.

Oxman, R. and Oxman, R. (1993). Precedents: Memory

structure in design case libraries. In CAAD Futures,

volume 93, pages 273–287.

Richter, K., Heylighen, A., and Donath, D. (2007). Looking

back to the future-an updated case base of case-based

design tools for architecture. Knowledge Modelling-

eCAADe.

Siebert, M. (2014). Concept of a CBR System for the Ar-

chitectural Conceptualization Phase Support. Project

report.

Voss, A. (1997). Case design specialists in FABEL. Issues

and Applications of Case-based Reasoning in Design,

pages 301–335.

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

156