Multi-Agent-Based Generation of Explanations for Retrieval Results
Within a Case-Based Support Framework for Architectural Design
Viktor Ayzenshtadt
1
, Christian Espinoza-Stapelfeld
1
, Christoph Langenhan
3
and Klaus-Dieter Althoff
1,2
1
University of Hildesheim, Institute of Computer Science, Samelsonplatz 1, 31141 Hildesheim, Germany
2
German Research Center for Artificial Intelligence (DFKI), Trippstadter Strasse 122, 67663 Kaiserslautern, Germany
3
Faculty of Architecture, Technical University of Munich, Arcisstrasse 21, 80333 Munich, Germany
Keywords:
Multi-Agent Systems, Case-Based Agents, Explanations, Architectural Design Support.
Abstract:
In this paper, we describe the general structure and evaluation of a multi-agent based system module that was
conceptualized to explain, and therefore, enrich the search results of the retrieval process within a distributed
case-based framework for support of early conceptual design phase in architecture. This explanation module
is implemented as an essential part of the framework and uses case-based agents, explanation ontology, and
explanation patterns as its underlying foundational components. The module’s main goal is to provide the user
with additional information about the search results to make the framework’s behavior during the retrieval
stage more transparent and traceable. System’s justification for displaying of results plays an important role
as well, and is also included in the explanations. We evaluated the explanation generation process with a
ground-truth set of explanations and a case-based validation process to ensure the suitability of the generated
explanation expressions for displaying in user interfaces connected to the framework. The results of the
evaluation confirmed our expectation and showed the general validity of the explanations.
1 INTRODUCTION
Computer-aided architectural design (CAAD) is a re-
search domain where one of its main tasks is to uti-
lize knowledge-intensive data about building designs
to support the early phases of architectural design
conceptualization. Case-based reasoning (CBR) is a
widely-used approach that can make use of such data
to provide designs similar to the currently created one.
Those similar designs are aimed to inform an architect
about possible similar semantic and relational struc-
tures of the prior designs found in a database (case
base), or show how the problem at hand was solved
in a different context. Other important objective of si-
milar designs (floor plans) is to provide the architect
with inspirational space to explore the designs that
can be helpful during the current project.
Many case-based design (CBD) approaches were
developed to utilize the methods of CAAD and CBR
to provide a tool for the support of the architectu-
ral conceptual design phase. MetisCBR (Ayzenshtadt
et al., 2016a) is one of such approaches that was cre-
ated as part of the basic research project Metis that
was aimed to combine the research areas of CAAD,
CBR, and multi-agent systems (MAS). MetisCBR
uses case-based agents and CBR-based retrieval met-
hods to find possibly helpful similar architectural de-
signs that correspond to the criteria entered by the
user. These criteria are based on a paradigm of se-
mantic fingerprint, that hierarchically represents meta
data and structural properties of a floor plan.
However, not only the information about the re-
trieved similar designs themselves is important for
speeding up and qualitative improvement of the ar-
chitectural design process. In many cases, users, in
our case: architects, want to have additional informa-
tion about the results: e.g., how exactly the system
achieved the given results, why exactly the results can
be helpful, or what are the structural differences of
query and result. That is, the users want the system to
be explanation-aware in order to provide explanations
for the returned search results.
In this paper, we present the recently developed
explanation module of MetisCBR, where case-based
agents work with explanation patterns to generate ex-
planations for the results returned during the retrieval
process. After the generation, the explanations are va-
lidated by a case-based comparison with an expert-
Ayzenshtadt, V., Espinoza-Stapelfeld, C., Langenhan, C. and Althoff, K-D.
Multi-Agent-Based Generation of Explanations for Retrieval Results Within a Case-Based Support Framework for Architectural Design.
DOI: 10.5220/0006650201030114
In Proceedings of the 10th International Conference on Agents and Artificial Intelligence (ICAART 2018) - Volume 1, pages 103-114
ISBN: 978-989-758-275-2
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
103
proven set of ground-truth explanations saved in a
special case base of explanations.
This paper is structured as follows: first, the inter-
nal MetisCBR-related work as well as external work
related to the purposes of the framework, and expla-
nations in CBR will be briefly presented. In the next
section, we describe MetisCBR more in detail, in-
cluding its underlying foundations for CBR-based re-
trieval and internal communication of agents. Next,
we in detail present the explanation module (Explai-
ner), including the general functionality, the charac-
terization and functionality of the agents, description
of explanation patterns and their use and modification
for the Explainer, and the underlying communication
pattern and its ontology. After that, the validation-
based evaluation of the generated explanations is pre-
sented. The conclusion section summarizes the pre-
sented work and gives a brief overview of the future
development of MetisCBR and the explanation appro-
ach.
2 RELATED WORK
In this section, we present work related to the purpo-
ses of MetisCBR, dividing it in external and internal
related work, as well as a short overview of work that
has been made so far in the domain of CBR-based ex-
planations.
2.1 Internal Related Work
During the basic research project Metis (KSD Re-
search Group, 2015) that aimed at combination of
MAS, CBR, and CAAD for architectural design sup-
port purposes, we developed a number of approa-
ches to reach this research goal. MetisCBR is one
of these developed approaches; a web-based floor
plan editor Metis WebUI (Bayer et al., 2015) that
can communicate with retrieval approaches, a retrie-
val coordination middleware KSD Coordinator (Roith
et al., 2016) that can query a number of (subgraph
matching-based) design retrieval engines, an android
app, and a touch-table-based application are other
examples. A comprehensive framework that unites
some of these techniques has been constructed as well
and presented in detail in (Sabri et al., 2017). Theo-
retical foundations developed during the project are
the semantic fingerprint (see also Section 3.2) and
the specification of the architectural graph description
language AGraphML (Langenhan, 2015).
2.2 External Related Work
A number of CBR-based approaches was introduced
in the past for the conceptual phase of the architec-
tural design process. In (Heylighen and Neucker-
mans, 2001) and (Richter et al., 2007), the correspon-
ding authors provide surveys of these approaches. In
(Richter, 2011), a comprehensive research into CBR
in architecture/CAAD was conducted to provide a
view on present, past, and future of combination of
these research areas. In (Richter, 2013) a summary of
this work is available that includes an overview of the
current problems of CBR in architecture.
Notable approaches are FABEL (Voss, 1997),
PRECEDENTS (Oxman and Oxman, 1993), SEED
(Flemming, 1994), DIM (Lai, 2005), or VAT (Visual
Architectural Topology, a semantic representation ap-
proach) (Lin, 2013). An approach that is most likely
closely related to the purpose of this paper is Case-
Book (Inanc, 2000), where a similarity explanation
report is mentioned as one of the system’s functio-
nalities. However, not much informaion is provided
about its mode of operation or evaluation.
The most notable approach from the research area
of multi-agent systems is (Anumba et al., 2007).
In this comprehensive work, examples and technical
foundations of use of MAS in construction and archi-
tecture domains are examined.
2.3 Explanations in Case-based
Reasoning
Explainability of case-based systems is one of the
questions of the CBR research field that has gai-
ned much attention during the last decades, produ-
cing a number of theoretical foundations and techni-
cal terms. Aamodt, one of the founders of the CBR’s
4R cycle (Retrieve, Reuse, Revise, Retain), was one
of the first authors to examine explanations for CBR
systems in (Aamodt, 1993). Roth-Berghofer descri-
bed foundational issues of explanations in CBR in
(Roth-Berghofer, 2004). Later, an explanation-aware
plugin myEACBR for the CBR prototypes framework
myCBR (Bach and Althoff, 2012) was developed in
(Lillehaug, 2011). Finally, in (Cassens and Kofod-
Petersen, 2007), application of explanation patterns to
CBR-based systems or intelligent systems in general
was examined.
We consider this previous work a valuable con-
tribution to the CBR research area and explanation-
aware systems. Our goal, therefore, was to take the
above described research as a basis and adapt it for
our specific purpose, the distributed case-based sup-
port for the architectural conceptualization phase.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
104
More specifically, the pattern approach described
in (Cassens and Kofod-Petersen, 2007) was conside-
red very similar to semantic fingerprint of architecture
(see Section 3.2) used by MetisCBR for retrieval, as
semantic fingerprints can be considered retrieval pat-
terns as well, which is documented in (Sabri et al.,
2017). This led us to adaptation of the explanation
patterns from (Cassens and Kofod-Petersen, 2007) for
our explanation module. That is, our aim in this case
was to achieve a sufficient degree of consistency be-
tween the system modules, using an established met-
hod and workflow for new agents. However, some
modifications during the application process of expla-
nation patterns were required, especially for detection
of patterns, as tasks of retrieval and explanation differ
in their nature (see Sections 4.2.1 and 4.2.2).
3 MetisCBR
MetisCBR is a multi-agent system that makes use of
case-based agents and CBR-based retrieval to find
possibly helpful previous solutions during the early
phases of architectural conceptual design. The frame-
work’s main aim is to efficiently improve the design
process in its early stages by providing the architects
with a tool that can find designs suitable for impro-
vement of the project at hand. The detailed descrip-
tions of the framework’s system architecture and the
domain model were already provided in (Ayzenshtadt
et al., 2015) and (Ayzenshtadt et al., 2016a). The
framework has been subject of different comparative
performance and qualitative evaluations with partici-
pation of other search methods, examples are (Ay-
zenshtadt et al., 2016b) and (Sabri et al., 2017).
The decision to design MetisCBR as a multi-agent
based software has a number of important reasons.
First, a system with agents that can communicate with
each other by means of applying a common language
and ontology allows us to unify the overall commu-
nication of the system modules. This is relevant, as
the queries and results of architectural designs used
by the system use a strict specification for their con-
struction and thus can be converted to an ontological
object for application in multi-agent frameworks. Se-
cond, some system components require to be able to
conduct a number of parallel operations, that should
not always be executed permanently, but, for example,
periodically. Agents design in many multi-agent fra-
meworks allows for execution of behaviors of diffe-
rent types concurrently – this fact was also important
for our choice of the system design paradigm. Third, a
requirement for each retrieval system within the Me-
tis project was to implement a coordination compo-
nent that is able to select a proper retrieval method or
strategy for the current query. Coordination of system
processes has a long tradition in the research field of
multi-agent systems, therefore, this fact played an es-
sential role in our decision as well. As described in
(Ayzenshtadt et al., 2015), we apply adapted methods
of coordination provided in (Nwana et al., 1996).
Further in this section, we provide a brief descrip-
tion of the general functionality of MetisCBR, and
the underlying concepts of semantic search patterns
(semantic fingerprints) and communications patterns.
Both of the pattern types are relevant for the purpose
of this paper.
3.1 General Functionality
The general functionality of MetisCBR is based
around the so-called retrieval containers that work
concurrently on resolving of search requests. In every
container, a number of CBR retrieval agents, that re-
trieve the case base of floor plans in order to find si-
milarly structured designs according to the selected
retrieval strateg, and a Manager, that controls the re-
trieval agents and manages the resolving of the cur-
rent query, are available. The search request itself
can consist of different sub-requests, where each of
these sub-requests represents a semantic fingerprint
(see Section 3.2) selected by the user (in this case,
a retrieval container is created for each sub-request,
the particular results of the containers are then amal-
gamated into a common result). On the higher level
of the system’s hierarchy, the Coordinator agent go-
verns the system’s processes, e.g., it selects the pro-
per retrieval strategy for the current search request (or
a sub-request) and assigns it to the corresponding re-
trieval container. The Coordination Team is a group
of Coordinator’s helper agents that consists of agents
for specific tasks, such as result collection from the
active retrieval containers. The agents communicate
by means of applying the specific communication pat-
terns that allow for standardization and unification of
processes within the system (see Section 3.3). The
general system structure with inclusion of the expla-
nation component is depicted in Figure 1.
3.2 Semantic Fingerprint
Semantic fingerprint of architecture (Langenhan and
Petzold, 2010) provides a description paradigm for
representation of architectural designs in CAAD sy-
stems. The concept of semantic fingerprint is aimed
at enrichment of the BIM (Building Information Mo-
deling) objects with semantic and topological infor-
mation contained in graphs derived from a floor plan.
Multi-Agent-Based Generation of Explanations for Retrieval Results Within a Case-Based Support Framework for Architectural Design
105
Web UI
CBR domain model
(myCBR API)
Neo4j
Search Request
Case
Bases
Coordinator +
coordination team
CBR agent 1
CBR agent 2
CBR agent 3
Room µ
Edge µ
Maintainer agent
GraphML agent
Imported
Cases
(Floor Plans)
Floorplan µ
Gateway
CBR manager
More retrieval containers ...
Retrieval container
Results
Explanation Deliverer
Explanation Creator
Case Base
of Expl.
Explainer
Expl.
Generation
Expl. Patterns
Figure 1: General structure of MetisCBR (with explanation module, the Explainer, included).
Fingerprint Name / Specics Fingerprint Name / Specics
FP1 Room Count
No connections between
rooms and no labels specied
FP5 Adjacency
Rooms information is
complete, no edge labels
FP2 Relation Count
No room information
specied
FP6 Accessibility
Edge information is complete,
no room labels
FP3 Room Graph
Anonymous representation
(no labels) of rooms & edges
FP7 Full Graph
All information about rooms
and edges available
FP4 Room Types
No room connections, only
room labels are specied
FP8 Natural Light
Light condition attributes
Figure 2: Semantic fingerprints currently implemented in MetisCBR.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
106
Therefore, each fingerprint is a graph or a subgraph
that can contain certain semantic and/or topological
information about the structure, e.g., room geometry,
of an architectural design. Semantic fingerprint fol-
lows an index-based hierarchical structure with Levels
that are divided in Units which contain certain Zones
with a number of Rooms. For retrieval purposes, se-
mantic fingerprints can be used as search patterns. In
this way, they are implemented in MetisCBR, and can
be specified by the user during the retrieval session. A
list of fingerprints (FPs) currently in use in MetisCBR
is shown in Figure 2.
3.3 Communication Patterns
The communication patterns of MetisCBR are the in-
ternal communication workflows that the agents make
use of during execution of their tasks in order to col-
laborate with each other. The communication pat-
terns were previously discussed in (Ayzenshtadt et al.,
2016c), in this section, similarly to the general struc-
ture of MetisCBR and semantic fingerprint, we give
only a brief description of the concept.
A communication pattern is a directed graph for
one of the system workflows, e.g., retrieval, where
nodes represent the agents involved in this workflow,
and edges represent steps or actions that are prede-
fined and either assigned by an agent of managing
type, such as Coordinator or Manager, or requested
by agents of another type. For each pattern, a sub-
set of the system’s communication ontology is used
to ensure the standardized communication among the
agents. The ontology, and thus a subset as well, con-
sists of three layers: object: abstract concepts of
query and result, data: concrete architectural data
from query or result, action: an action that should be
executed with this data. Therefore, the pattern steps
are named after the ontology concept (action class)
that holds information about this type of action. In
Figure 3, the communication pattern ‘retrieval’ is de-
picted for demonstration of a pattern structure.
4 EXPLANATION MODULE
(EXPLAINER)
Users of (decision) support systems are likely to in-
crease their satisfaction with the system if the system
is able to produce and present reasonable traces of its
behavior. Therefore, our motivation behind the deve-
lopment of the explanation component is to give our
user group, the architects, a sufficient amount of in-
sights into how the system achieved the results contai-
ned in the final result list. In this section, we in detail
present the foundations of the explanation module Ex-
plainer: agents, explanation patterns, communication
pattern, and construction and validation of explana-
tion expressions.
4.1 Explainer Agents
As depicted in Figure 1, the general structure of the
Explainer employs two agents that are responsible for
construction, validation and transfer of explanations:
the Explanation Deliverer and the Explanation Crea-
tor.
4.1.1 Explanation Deliverer
The Deliverer agent of the explanation module is re-
sponsible for receiving of explanation requests from
the managers of the active retrieval containers and
sending back the corresponding explained results to
these managers. After receiving an explanation re-
quest, that at first consists of the query that has to be
explained, the Deliverer saves this query in a speci-
fic explanation schedule and waits for retrieval results
to be completed. When the resolving of the query is
finished by the corresponding retrieval container, the
manager of this container sends the results to the De-
liverer. The Deliverer then retrieves the query from
the schedule, constructs an object that consists of the
query and the corresponding result set, a so-called
query-result object, and sends it to the other agent of
the explanation module, the Explanation Creator. Af-
ter receiving the explained results from the Creator,
the Deliverer forwards them to the corresponding ma-
nager who in turn sends them to the Result Collector
agent of the Coordination Team (previously mentio-
ned in Section 3, see more about Result Collector’s
role in 4.1.3).
4.1.2 Explanation Creator
The Explanation Creator agent is responsible for the
creation/construction of explanations for the result set
of a query. After receiving the query-result object
from the Deliverer, the Creator tries to generate an ex-
planation for each result in the result set of this object,
based on the semantic fingerprint of the query. The
explanation is then added to the corresponding result
in the result set. At this point, the query-result object
contains the explained results and is sent back to the
Deliverer. The Creator’s mode of operation during the
actual explanation generation process consists of two
subsequent phases:
1. Detection of explanation patterns (see Section
4.2.1) and creation of explanation texts with pre-
defined expressions (see Section 4.2.1)
Multi-Agent-Based Generation of Explanations for Retrieval Results Within a Case-Based Support Framework for Architectural Design
107
GraphML
CBR
Manager
Coord.
Sub-
Coord.
CBR
Retrieval
CBR
Manager
Coord.
Coord.
Parse
Forward
Forward
X
H
R
Gateway
F
o
r
w
a
r
d
S
a
v
e
S
k
i
p
T
i
m
e
o
u
t
S
o
l
v
e
Solve
Fo
rw
a
rd
S
a
v
e
Sub-
Coord.
F
o
r
w
a
r
d
Query
Result
Result
Result
Queries
Timeout
Sub-Pattern
Sub-Pattern
Figure 3: The communication pattern ‘retrieval’ (including the sub-pattern for retention of the currently entered query).
2. Case-based validation of explanations (see
Section 4.3)
4.1.3 Other Agents
A special role in the explanation process is assigned
to the Result Collector agent, despite the fact that
this agent does not belong to the explanation module
agents. The Result Collector receives the set of ex-
plained results for each fingerprint (i.e., sub-request)
of the query and, while amalgamating the fingerprint
results (i.e., computing the common similarity value
for the complete result), it selects the most valid ex-
planation (see Section 4.3) among all available expla-
nations and adds it to the result information in the fi-
nal result set that is then presented to the user.
Another important role is played by the Manager
agent of a retrieval container, who is, among its ot-
her functions, the communication interface between
the Explainer and the Result Collector. However, the
Manager does not apply any modifications to the re-
sult data (where the explanations are included).
4.2 Explanation Patterns
Explanation patterns for case-based reasoning appli-
cations and approaches is a concept developed in
(Cassens and Kofod-Petersen, 2007). In the narro-
wer sense, the explanation patterns are also known
as Explanation Problem Frames and extend the ori-
ginal concept of problem frames presented by Jack-
son in (Jackson, 1999). Taking the Jackson’s problem
frames as a basis for description of problems in soft-
ware engineering, the explanation problem frames for
intelligent systems were contructed in (Cassens and
Kofod-Petersen, 2007) and specified a number of spe-
cial patterns that serve as foundation for creation or
generation of explanations in such systems.
Following the structure proposed by Jackson, the
explanation problem frames consist of a computation
unit machine, representation of a real-world domain,
and the reference to the requirements for a problem
that has to be solved. All three components of a pro-
blem frame should be connected with each other in
order to provide a common foundation for creation
of an explanation. In (Cassens and Kofod-Petersen,
2007), a number of explanation problem frames were
identified for CBR-based systems, that are aimed at
achieving, inter alia, the following, expert systems-
inspired, explanation goals:
Justification Explanation of why the given ans-
wer or solution is good at this point of wor-
king/interacting with the system. The goals is
to make the user confident with the system’s
decision-making by reasonably justifying the ans-
wer as suitable and helpful for the current situa-
tion, question, or query.
Transparency Explanation of how the system
found the given answer or solution. The goal is to
give the user some insight into how the system’s
reasoning works, that is, make the decision-
making processes of the system more transparent.
Relevance – Explanation of why the question that
the system is asking is relevant in the current si-
tuation. For example, if an insufficient amount of
properties or attributes of the failure was provided
by the user to find a proper solution, the system
may ask for further refinement of the problem des-
cription, thus ask the user for more relevant data.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
108
However, the system should be able to provide a
reasonable explanation of why this refinement is
relevant, i.e., in this example, why the data provi-
ded by the user is insufficient.
4.2.1 Modification for MetisCBR
In a recent work in (Espinoza, 2017), the general con-
cept of explanation patterns was extended for use in
MetisCBR’s results explanation module. For this pur-
pose, the explanation patterns described in Section
4.2.1 were modified to work with the underlying con-
cepts of MetisCBR, such as semantic fingerprint and
the domain model of the case base of architectural
designs. Based on the machine/domain/requirement
structure of the problem frames, the explanation
patterns were adapted for the domain of semantic
fingerprint-based CBR-retrieval of architectural de-
signs, taking the fingerprint context (more precisely,
constraint) as its basis.
The explanation patterns for MetisCBR were con-
ceptualized w.r.t explanation goals, previously des-
cribed in 4.2.1. That is, three explanation patterns
were created to extend MetisCBR with the explaina-
bility function. For each pattern, the fingerprint con-
straint serves as a basis for the problem frame struc-
ture, it is then integrated with justification, transpa-
rency, and relevance frames. In Figure 4 the integra-
tion of patterns for justification and transparency is
shown: a new machine Fingerprint has been added to
the previously existing Justification and Transparency
to allow for connection of these original patterns to
the main properties (meta data, rooms, edges) of the
fingerprint- and CBR-based retrieval of floor plans.
4.2.2 Detection of Patterns
The explanation patterns were implemented in Metis-
CBR in the way, that the system (more precisely, the
Explanation Creator agent) tries to detect if a pattern
is available for the current query and the current re-
sult. Only if the pattern is available, the Creator will
add the corresponding explanation expression to the
explanation text string of the result.
The detection conditions are saved in a special
heuristic ruleset and are different for each pattern and
fingerprint. That is, it depends on complexity and type
of the fingerprint (FP) if the pattern can be detected.
Below, some examples are provided to demonstrate
heuristics of rules of pattern detection:
FP 1, 2, 4, 8: The pattern ‘Transparency’ is
detected if 2/3 properties from {Room Count,
Edge Count, Room Types} could be detected in
the meta data of the query floor plan.
Fingerprint
machine
Floor plan
meta data
Room
Edge
Justication
machine
Information
display
Specic case
knowledge
Reasoning
trace
Justication
explanation
Transparency
explanation
Transparency
machine
Display
Fingerprint-
based
explanation
X
C
C
C
X
X
X
X
Figure 4: Modification of explanation patterns for Metis-
CBR. Enclosed in blue are the original problem frames from
(Cassens and Kofod-Petersen, 2007). C stands for the goal
of explanation, X denotes the system knowledge. The In-
formation Display goal is mandatory for connection to the
user interface.
FP 1, 2, 4, 8: The pattern ‘Justification’ is
detected if the similarity grade of the result floor
plan is better than unsimilar.
FP 6: The pattern ‘Transparency’ is detected if
all properties from {Room Count, Edge Count,
Edge Types} could be detected in the meta data
of the query floor plan.
FP 7: The pattern ‘Justification’ is detected
if the similarity grade of the result floor
plan is better than unsimilar and all pro-
perties from {Room Count, Edge Count, Room
Types, Edge Types} could be detected in the
meta data of the query floor plan.
As seen in the examples provided above, the simi-
larity grade plays an important role in the detection
of the explanation patterns. This similarity grade de-
pends on the similarity value of the result floor plan
and is classified as follows: as very similar we clas-
sify the result if its similarity Sim 0.75, similar if
0.75 > Sim 0.5, sufficiently similar if 0.5 > Sim
0.25, and unsimilar if Sim < 0.25.
4.2.3 Construction of Explanations
The actual explanation is a combination of explana-
tion expressions as text strings, where for each of the
patterns for each of the fingerprints and for each of
Multi-Agent-Based Generation of Explanations for Retrieval Results Within a Case-Based Support Framework for Architectural Design
109
the similarity grades a short and long (full) version
of an expression exist. An explanation text is then
a concatenation of particular expressions, it is possi-
ble to combine both short and full versions of an ex-
pression, however, either short or full version of the
same category (similarity grade, fingerprint, or pat-
tern) should be used. In Figure 5, examples of the
explanation texts are shown for an exemplary query
and the corresponding result.
All text strings for full explanations were provided
by an expert from the architectural design domain to
ensure the suitability and understandability of expla-
nations for the representatives of this domain.
4.3 Validation of Explanations
The aim of validation of explanations is to prove if the
generated explanation corresponds to the architectural
understanding and technical term requirements. We
consider the validation an important step of the pro-
cess, as it helps to increase the possibility of returning
of more usergroup-friendly explanation expressions.
To validate an explanation we decided to apply a
case-based validation approach. That is, an explana-
tion is validated against a ground-truth set of expla-
nations saved in a case base of explanations. Each
explanation in this set consists of a number of attribu-
tes (see Table 1), most of those attributes are used for
comparison with their counterparts from the genera-
ted explanations. The unused attributes are used for
information purposes.
After the attribute-wise comparison, where for
each relevant attribute a similarity value is computed,
these attribute similarity values are amalgamated with
a weighted sum (see (1)) to produce a validation simi-
larity value v, that shows how similar the query expla-
nation is to the currently compared ground-truth ex-
planation. The weight values for the attribute simila-
rities used in this weighted sum depend on the number
of detected explanation patterns. For example, if two
patterns were detected, the weight value ω
t
for text si-
milarity t is 0.2, the weight ω
p
for a pattern similarity
p is 0.4 (if only one pattern was detected: 0.3/0.7 ac-
cordingly). This dynamic distribution of weights was
applied in order to avoid penalization of results where
not all of the patterns could be detected, but a highly
valid explanation can be expected anyway. That is,
we are interested in results with most reasonable ex-
planations possible.
The highest validation similarity that indicates the
most similar explanation from the set of all validated
similarity values V then becomes an explanation simi-
larity of the query similarity, i.e., the currently gene-
rated explanation. The generated explanation is valid
if its explanation similarity exceeds a threshold value,
e.g., 0.5.
v = ω
t
t + ω
p
n
i=1
p
n
, v V
(1)
4.4 Explainer Communication and
Ontology
To exchange information among each other, the
agents of the Explainer make use of the communica-
tion patterns described in 3.3. That is, they follow the
workflow communication steps defined in the pattern.
For the Explainer, a specific communication pattern
‘explainer’ was created that underlies all communi-
cation processes within the explanation module. Si-
milarly to the ‘retrieval’ pattern shown in Figure 3,
where Query and Result objects hold the concrete in-
formation about the floor plans from the correspon-
ding query and result, the ‘explainer’ pattern contains
the Explain object that contains information about
data that should be enriched with explanations. In this
case, this is the previously mentioned query-result ob-
ject that includes query and result data before, and is
enhanced with explanation expressions for each result
after the explanation generation process. The commu-
nication pattern ‘explainer’ is shown in Figure 6.
The ‘explainer’ communication pattern relies on
the explainer ontology that contains all possible con-
cepts that can occur during collaboration between
both agents of the Explainer (Deliverer and Creator)
and agents of the retrieval module (Manager, Result
Collector). A special concept is Explanation, that
consists of data for explanation validation as descri-
bed in Section 4.3. All ontological concepts are re-
presented as objects in FIPA-SL content language.
5 EVALUATION
To evaluate the current functionality of the Explainer,
we decided to conduct a comprehensive case-based
validation of generated explanations by means of que-
rying MetisCBR with a number of different queries.
The process of case-based validation of an explana-
tion is described in Section 4.3. In this section, we
first give a brief description of the setting of the expe-
riment, followed by the results achieved during the
evaluation. To our knowledge, no comparable ex-
planation tool in current architectural design support
software exists at the moment, i.e., this is the first eva-
luation of a system module of such kind. Our general
expectation was, that the Explainer can produce a suf-
ficient number of queries that are valid for inclusion
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
110






Justification
Pattern text (full)
Transparency
Pattern text
(short)
!
!
"#!
$% 
Justification
Pattern text (full)
Transparency
Pattern text
(short)
&'(
)

'
)* 
Relevance
Pattern text (full)
Transparency
Pattern text
(short)
+
,
-,
.
+
Figure 5: An examle of constructed explanations between an exemplary qurey and the corresponsing result. Bubbles denote
rooms without specific functionality and/or geometry. C denotes the CORRIDOR room type, others are LIVING, SLEEPING,
PARKING, and KITCHEN. Fingerprints (FPs) are described in Figure 2.
Expl.
Creator
E
x
p
l
a
i
n
Expl.
Deliverer
Query
Result
Result
Result
+
Manager
Explain
Q
u
e
r
y
-
R
e
s
u
l
t
o
b
j
e
c
t
Query
F
o
r
w
a
r
d
Q
u
e
r
y
+
e
x
p
l
a
i
n
e
d
r
e
s
u
l
t
s
Explained results
Forward
Result
Collect.
Generation of
explanations
Figure 6: The communication pattern ‘explainer’(the order of actions is represented clockwise, starting at Manager).
Table 1: Attributes of the Explanation concept in the case base of explanations.
Attribute Type Description Similarity Function
id string Internal Explanation ID not in use
Case string Reference to the case (result) not in use
Query string Reference to the query not in use
Text string Text of the explanation Levenshtein distance-based sim.
PatternJustification boolean Justification pattern available? boolean comparison
PatternRelevance boolean Relevance pattern available? boolean comparison
PatternTransparency boolean Transparency pattern available? boolean comparison
Multi-Agent-Based Generation of Explanations for Retrieval Results Within a Case-Based Support Framework for Architectural Design
111
K
K
K
W
L
L SL
S
S
S
B
B
S
L
SL
C
Figure 7: Examples of evaluation queries. Bubbles denote rooms without specific functionality and/or geometry. C denotes
the CORRIDOR room type, others are WORKING, LIVING, STORAGE, SLEEPING, BATH, PARKING, and KITCHEN.
into the final result set, thus for showing them to the
user group of the framework.
5.1 Setting
We performed the evaluation on a set of 225 retrieva-
ble architectural designs (floor plans) saved as cases
in the case base of MetisCBR’s retrieval component.
The maximum complexity of a floor plan was 368,
i.e., the maximum value of count c C
R
C
E
, where
C
R
is the set of all room count values and C
E
is the set
of all edge counts. To provide a ground-truth set for
the evaluation, a case base of 15 explanations (with
case structure as described in Section 5) was crea-
ted. This ground-truth set then became a validation
base for the explanations generated during the expe-
riment. Each explanation case in this base was manu-
ally examined to be a valid explanation itself – in this
case, this means it has been checked for the technical
validity of the combination query/result/similarity va-
lue/explanation. It has also been checked that the dif-
ferences between the fingerprints and corresponding
explanations are sufficient to provide the system with
space for validation comparison, i.e., the FPs were se-
lected in the way that they do not appear too often
and the corresponding explanation expressions were
all from different similarity grades.
As queries, 30 floor plans of different complex-
ity and abstraction level were used. These floor plans
were created with one of MetisCBR’s floor plan edi-
tors, the Metis WebUI. In Figure 7, some examples of
the query floor plans are shown. On average, 2 FPs
were used per query, making 59 sub-requests in total.
The FPs were selected in the similar way as in the ex-
planations ground-truth set, i.e., they were prevented
from appearing too often.
5.2 Results
The results revealed that our expectation about the ge-
neral suitability of the explanation module has been
met. With a total number of 67875 constructed ex-
planations, we built a sufficient basis for evaluation
of the Explainer’s current capabilities. From this total
explanation count, 57575 (84.825%) were valid and
10300 (15.175%) were invalid, this confirmed our as-
sumption that the case-based validation is a reasona-
ble method to validate such an explanation. Regard-
less of relatively high threshold value of 0.5 for vali-
dity rating, the system could produce a high amount
of valid explanations that can be shown to the user. In
the following sections, some detailed characteristics
and findings from the experiment are presented.
5.2.1 Detected Patterns
The number of the explanation patterns detected for
the retrieval results varied highly between the parti-
cular patterns. Relevance was detected in 1710 cases,
whereas Transparency in all 67875 cases, and Justifi-
cation in 50850 cases. However, these are expectable
values, considering the fact that the system does not
always require additional information from its user,
i.e., does not provide questions to the user.
In Figure 9, the distribution of combinations of
patterns is shown. Here, the expected high value for
Justification/Transparency and low numbers of com-
binations with Relevance can also be seen.
JT JTR TR
49,290
1,560
150
Pattern Combinations
Amount of Combinations
Figure 8: The distribution of explanation pattern combinati-
ons detected in the results. J, T, and R stand for Justification,
Transparency, and Relevance respectively.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
112
5.2.2 Similarity Distribution
As established for the experiments in our work, we
show how the similarity grades (see Section 4.3) were
distributed among the retrieval results (Figure 10), ba-
sed on total count of 4525 results. Firstly, however,
we show the difference between average validation
similarity and average results similarity (Figure 9).
Avg Expl. Sim. Avg Result Sim.
0
0.2
0.4
0.6
0.8
1
0.78
0.44
Average Sim. Values
Sim. Scale
Figure 9: The average similarity values for results and ex-
planations.
Unsim. Suff. sim. Sim. Very sim.
1,126
1,675
202
1,522
Similarity Grades
Results count
Figure 10: The distribution of results counts among the si-
milarity grades.
6 CONCLUSION
In this work, we presented the structure and main
components of the explanation module (the Explai-
ner) for retrieval results of MetisCBR for case-based
search of similar architectural designs. The module
employs two agents, communication pattern and on-
tology, as well as explanation patterns that build the
theoretical base for creation of explanation expressi-
ons.
The generated explanations are validated with a
case-based validation by comparison with ground-
truth explanations saved in a case base of explana-
tions. The validation-based evaluation of generated
explanations showed their general suitability for pre-
sentation to the user group as part of information in
the retrieval results.
Our future work on the Explainer will be con-
centrated on its further development with a goal to
built more fine-grained explanation patterns and de-
tection rulesets, that will be based not only on seman-
tic fingerprints but on the attributes of rooms, relati-
ons, and the meta data of the design graph. To eva-
luate these patterns, a comprehensive study with par-
ticipation of architectural domain representative will
be conducted.
REFERENCES
Aamodt, A. (1993). Explanation-driven case-based reaso-
ning. In European Workshop on Case-Based Reaso-
ning, pages 274–288. Springer.
Anumba, C., Ren, Z., and Ugwu, O. (2007). Agents and
multi-agent systems in construction. Routledge.
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 Reasoning
(UKCBR-2015), located at SGAI International Con-
ference on Artificial Intelligence, December 15-17,
Cambridge, United Kingdom. School of Computing,
Engineering and Mathematics, University of Brigh-
ton, UK.
Ayzenshtadt, V., Langenhan, C., Bukhari, S. S., Althoff, K.-
D., Petzold, F., and Dengel, A. (2016a). Thinking with
containers: A multi-agent retrieval approach for the
case-based semantic search of architectural designs.
In Filipe, J. and van den Herik, J., editors, 8th Inter-
national Conference on Agents and Artificial Intelli-
gence (ICAART-2016), February 24-26, Rome, Italy.
SCITEPRESS.
Ayzenshtadt, V., Langenhan, C., Roith, J., Bukhari, S. S.,
Althoff, K.-D., Petzold, F., and Dengel, A. (2016b).
Comparative evaluation of rule-based and case-based
retrieval coordination for search of architectural buil-
ding designs. In Goel, A., Roth-Berghofer, T., and
Diaz-Agudo, B., editors, Case-based Reasoning in
Research and Development. 24th International Con-
ference on Case Based Reasoning (ICCBR-2016), Oc-
tober 31 - November 2, Atlanta, Georgia, USA. Sprin-
ger, Berlin, Heidelberg.
Ayzenshtadt, V., Mikyas, A., Althoff, K.-D., Bukhari, S. S.,
and Dengel, A. (2016c). Ontology-based communi-
cation architecture within a distributed case-based re-
trieval system for architectural designs. In Mueller,
E., Krestel, R., and Radtke, G., editors, LWDA 2016
Multi-Agent-Based Generation of Explanations for Retrieval Results Within a Case-Based Support Framework for Architectural Design
113
- Lernen, Wissen, Daten, Analysen - Workshop Pro-
ceedings. GI-Workshop-Tage ”Lernen, Wissen, Daten,
Analysen” (LWDA-2016), September 12-14, Potsdam,
Germany. Hasso-Plattner-Institut.
Bach, K. and Althoff, K.-D. (2012). Developing case-based
reasoning applications using myCBR 3. In Case-
Based Reasoning Research and Development, pages
17–31. Springer.
Bayer, J., Bukhari, S., Dengel, A., Langenhan, C., Alt-
hoff, K.-D., Petzold, F., and Eichenberger-Liwicki,
M. (2015). Migrating the classical pen-and-paper ba-
sed conceptual sketching of architecture plans towards
computer tools - prototype design and evaluation. 11th
IAPR International Workshop on Graphics Recogni-
tion - GREC’15, Nancy, France.
Cassens, J. and Kofod-Petersen, A. (2007). Designing ex-
planation aware systems: The quest for explanation
patterns. In ExaCt, pages 20–27.
Espinoza, C. (2017). Analysis of Identification of Explana-
tion Patterns for an Explanation Module for Support
of Design Phase in Architectural Domain. Project re-
port. University of Hildesheim.
Flemming, U. (1994). Case-based design in the SEED sy-
stem. Automation in Construction, 3(2):123–133.
Heylighen, A. and Neuckermans, H. (2001). A case base of
case-based design tools for architecture. Computer-
Aided Design, 33(14):1111–1122.
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.
Jackson, M. (1999). Problem analysis using small problem
frames. South African Computer Journal, pages 47–
60.
KSD Research Group (2015). KSD - Forschungsprojekte -
metis. http://ksd.ai.ar.tum.de/?page_id=140
Last visited: 19.11.2017.
Lai, I.-C. (2005). Dynamic idea maps: a framework for lin-
king ideas with cases during brainstorming. Interna-
tional journal of architectural computing, 3(4):429–
447.
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 Mo-
dern Information Technologies, 4:13.
Lillehaug, M. B. (2011). Explanation-aware case-based re-
asoning. Master’s thesis, Institutt for datateknikk og
informasjonsvitenskap.
Lin, C.-J. (2013). Visual architectural topology. In Open
Systems: Proceedings of the 18th International Con-
ference on Computer-Aided Architectural Design Re-
search in Asia, pages 3–12.
Nwana, H. S., Lee, L. C., and Jennings, N. R. (1996). Coor-
dination 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. (2011). Augmenting Designers’ Memory: Case-
based Reasoning in Architecture. Logos-Verlag.
Richter, K. (2013). What a shame-why good ideas can’t
make it in architecture: A contemporary approach to-
wards the case-based reasoning paradigm in architec-
ture. In FLAIRS Conference.
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.
Roith, J., Langenhan, C., and Petzold, F. (2016). Supporting
the building design process with graph-based methods
using centrally coordinated federated databases. In
16th International Conference on Computing in Civil
and Building Engineering.
Roth-Berghofer, T. R. (2004). Explanations and case-based
reasoning: Foundational issues. Advances in case-
based reasoning, pages 195–209.
Sabri, Q. U., Bayer, J., Ayzenshtadt, V., Bukhari, S. S., Alt-
hoff, K.-D., and Dengel, A. (2017). Semantic pattern-
based retrieval of architectural floor plans with case-
based and graph-based searching techniques and their
evaluation and visualization. In ICPRAM, pages 50–
60.
Voss, A. (1997). Case design specialists in FABEL. Issues
and Applications of Case-based Reasoning in Design,
pages 301–335.
ICAART 2018 - 10th International Conference on Agents and Artificial Intelligence
114