TOWARDS MULTIAGENT INTERACTIVE ADAPTION

OF INDIVIDUAL MODELS FOR DECISION SUPPORT

∗

Tobias Warden, Thomas Wagner, Hagen Langer and Otthein Herzog

Center for Computing and Communication Technologies (tzi), Universität Bremen, Bremen, Germany

Keywords:

Multiagent systems, Interaction protocols, Distributed knowledge management, Machine learning, Knowl-

edge exchange, Interactive learning.

Abstract:

Software agents in complex, dynamic environments need to update, adapt, and improve their knowledge mod-

els for decision making in order to achieve adequate results. Their individual adaption often relies on machine

learning from observational data. However, when data is not available in the required quantity and quality,

alternative approaches are required. We propose an interaction-based approach to individual model adaption

in multiagent systems, describe agent roles and interaction principles, and discuss how a goal-oriented transfer

of knowledge among agents can be integrated into an agent-based knowledge management framework.

1 INTRODUCTION

Besides adequate access to decision-relevant informa-

tion and domain knowledge, access to tailored dy-

namic knowledge is a competitive advantage for the

performance of intelligent agents in complex task do-

mains such as health care, supply chain management,

electronic markets, or autonomous logistics. Refer-

ring to dynamic knowledge, we concur with Jennex

who deﬁnes knowledge as information, ”understood

such that it explains the how and the why about some-

thing” (Jennex, 2009), i.e., patterns applicable in sit-

uation assessment and prediction. Dynamic knowl-

edge subsumes several categories of integral knowl-

edge models for decision making, including decision

models, prediction models, or classiﬁcation models.

Depending on the speciﬁc use case, the respective

models may be placed at an agent’s disposal on in-

stantiation. This means that the models have been ei-

ther learned ofﬂine from historic data or handcrafted

in a knowledge engineering process. In other cases,

the agents themselves may be equipped with the pre-

requisites for online learning of models, based on

individual experience or accessible data repositories.

We consider agents which not only compile and em-

ploy classiﬁcation models in their primary domain

∗

This research has been supported by the German Re-

search Foundation (DFG) within the Collaborative Research

Centre 637 ”Autonomous Cooperating Logistic Processes

– A Paradigm Shift and its Limitations” at the Universität

Bremen, Germany.

roles but also possess the ability to assess the perfor-

mance of their models. This kind of meta-reasoning

in particular enables the agents to monitor their per-

formance over time and determine when an adaption

of momentarily operationalized classiﬁcation models

is necessary.

However, left to its own devices, it is often not

possible for a deployed agent to effectuate the desired

model adoption. For instance, the necessity for adap-

tion may allude to a preset model, provided for use

by an external source such as a knowledge engineer.

Consequently, the agent may not have access to a rep-

resentative pool of data to address adaption by means

of individual re-learning on this data. In other cases,

the ability to relearn may exist but the costs of a new

learning sweep are prohibitive or it has become ev-

ident that available learning schemes cannot provide

an improved model. Thus, while self-assessment of

an agent reveals a necessity to adapt a decision sup-

port model, its skill set and/or resources inhibit a self-

sufﬁcient adaption.

We claim that the multiagent environment pro-

vides the key to complement the individual adaption

capability of an agent. Within a persistent multiagent

system (MAS), there are often multiple agents with

kindred primary tasks and associated decision sup-

port models. These models may contain knowledge

that is both relevant and useful for an agent with an

adaption deﬁciency. We propose a knowledge man-

agement (KM) framework which allows agents to tap

that knowledge in a goal-oriented way. Our approach

553

Warden T., Wagner T., Langer H. and Herzog O..

TOWARDS MULTIAGENT INTERACTIVE ADAPTION OF INDIVIDUAL MODELS FOR DECISION SUPPORT.

DOI: 10.5220/0003886905530559

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (IWSI-2012), pages 553-559

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

builds upon dedicated knowledge management func-

tions executed by both the agent seeking model adap-

tion and its supporters. It enables individual adaption

in knowledge networks. Speciﬁcally, it acknowledges

that dynamic knowledge may not be simply trans-

ferred among agents. Rather, interactive adaption as

we see it requires that the agents seeking adaption ac-

tively interpret and integrate advice from peers, thus

effectively re-constructing individual advisor knowl-

edge. Thus, our approach to address adaption deﬁ-

ciencies through social interaction emulates, in spirit

rather than cognitive accuracy, human problem solv-

ing strategies. Also, instead of imposing a ﬁxed orga-

nizational structure, the approach is fully distributed.

The framework builds upon and extends approaches

pursued in the machine learning (ML) community,

such as active learning or transfer learning.

The following section ﬁrst provides a detailed

problem analysis and relates our methodology to ap-

proaches from the ML community. Section 3 then

adds a knowledge management and multiagent per-

spective. It introduces our framework that enables in-

teractive adaption of individual decision support mod-

els. Section 4 highlights related research, then con-

clude with a discussion and future work.

2 PROBLEM ANALYSIS

The minimal common prerequisites for agents that we

seek to support through interactive model adaption

are: 1) For competitive action in their primary domain

role, the agents rely on a classiﬁcation model. We

make no assumptions as to the origin of this model,

although models learned by the agent itself are our

primary concern. 2) The agents actively monitor both

performance of their models and any self-sufﬁcient

model adaption. Based on these prerequisites, a range

of scenarios for our approach can be spanned.

The ﬁrst distinction refers the embedding of a

learning subsystem. Agents without this feature can-

not adapt their model through learning from training

data. For adaption, they are thus necessarily depen-

dent on external help. For agents with a learning sub-

system, a second distinction refers to whether indi-

vidual learning constitutes a feasible option alterna-

tive in the momentary context. Engaging in learn-

ing may be infeasible due to the following reasons:

1) Training data may not be accessible, at all, or be

insufﬁcient. 2) The costs in terms of either com-

putation/memory resources or computation time in-

curred by (re-)learning a model may be prohibitive.

In both cases, self-sufﬁcient learning is not practi-

cal as means to improve a model. Hence, agents in

these situations have an interest in enlisting external

help for a target-oriented direct model modiﬁcation.

Even if (re-)learning was a feasible option in princi-

ple, agents may still have to face problems: The avail-

able data constitutes a skewed sampling of an under-

lying true distribution and is thus not representative

or data may only allow to learn a concept description

which does not comply with desired quality character-

istics. For classiﬁcation, this would be objectives with

respect to acceptable precision and recall, or false-

positive/false-negative classiﬁcations. While learning

is possible in these cases, due to the input data it does

not provide a model which complies with preset per-

formance standards or which bests the currently op-

erationalized model. Therefore, agents in such situ-

ations may have an interest in enlisting external help

designed to either broaden their data basis or acquire

advice which essentially allows their machine learn-

ing scheme to derive additional value from an unvar-

ied data basis.

We have shown scenarios where agents can draw

a surplus from a ﬂexible framework for interactive

adaption of their models. An additional important

ﬁnding is that the required complementary adaption

skill sets vary depending on the respective situation.

From a machine learner’s perspective, selected prob-

lem instances laid out in the problem analysis have

been addressed in the literature. Two strands of ML-

research with special relevance are transfer learning

from previous learning tasks an interactive learning

supported by other individuals.

Transfer learning addresses the speciﬁc problem

of an insufﬁcient data basis to learn a model through

re-use of training data originally collected for related

learning tasks (for a survey on transfer learning for

classiﬁcation problems, see (Pan and Yang, 2009)). A

broadening of the data basis is also a primary goal in

the area of active learning (Settles, 1995). The ap-

proach has been suggested amongst others for clas-

siﬁcation problems where the labeling of training in-

stances for learning is costly or time-consuming. Be-

ginning with a small pool of labeled training data,

the learning system iteratively learns a model and as-

sesses which additional data would provide the best

chance to optimize the hitherto learned model. This

data is then made available through intervention of a

human domain expert or acquired autonomously, e.g.,

through conduct of experiments.

Možina et al. have proposed argumentation based ma-

chine learning (ABML). This approach allows to at-

tain improvements in the performance of a learning

system which needs to solve a classiﬁcation task by

a human domain expert as an interaction counter-

part (Možina et al., 2007). The learning system is

ICAART 2012 - International Conference on Agents and Artificial Intelligence

554

endowed with self-assessment capabilities in that it

monitors its own learning progress and, speciﬁcally,

identiﬁes problem instances in the training data, that

are particularly ill-covered by its learned concept de-

scription. The learner then reaches out to its preset hu-

man interaction partner, presenting these instances as

queries. The expert then uses his domain knowledge

to provide a machine-readable explanation (called ar-

gumentation). These are accounted for in subsequent

learning phases.

ABML addresses what we consider the most chal-

lenging ﬂavor of interactive model adaption in that it

tries to tap on the implicit domain expertise of an ad-

visor to augment the existing advisee training data,

thus enabling learning progress. However, the ap-

proach needs to be elevated to multiagent learning.

Here, other learning agents as artiﬁcial domain ex-

perts would then subsume the single human expert

involved in ABML. Since speciﬁc strengths of ABML

that have been highlighted by Možina et al. include

not only ”effective reconstruction of expert’s prior

knowledge” (Možina et al., 2007, p. 932) and favor-

able results even in situations where the human in-

teraction partner’s ”knowledge about the domain is

superﬁcial (might be wrong occasionally)” (Možina

et al., 2007, p. 933), this approach seems promising.

3 A MULTIAGENT APPROACH

FOR INTERACTIVE ADAPTION

The methodical assessment of robust interactive adap-

tion and, in the process, transfer of individually

learned knowledge among cooperating agents in mul-

tiagent systems requires a suitable context from a

knowledge management point of view. We draw on

a framework for distributed knowledge management

that we have originally proposed for intelligent agents

jointly realizing control of autonomous logistic pro-

cesses (Langer et al., 2006). We will use and extend

this framework as a point of origin to derive necessary

knowledge management roles, means for interoper-

ability, and intra- as well as inter-agent organization

of multiagent adaption.

3.1 Role-based Distributed Knowledge

Management in MAS

This framework focusses on KM whose functions are

procured by software agents. More importantly, these

functions are also designed for use by agents as ar-

tiﬁcial decision-makers. In complex task environ-

ments, e.g., autonomous logistics, the availability of

diverse knowledge has been identiﬁed as a key factor

for an effective treatment of the primary agent roles in

the modeled processes with a desired quality of ser-

vice (Gehrke et al., 2010). Often, initial provision of

such default knowledge alone is not sufﬁcient in order

to accommodate for the complexity and dynamics of

the task environment. It then becomes necessary to

design adaptive agents, capable of individual knowl-

edge revision and the compilation of tailored models

via learning. Over time, knowledge hence becomes

to an increasing degree tailored to its task context.

Thus, analogously to the situation with employees

within organizations, knowledge is spread rather than

accumulated in a centralized knowledge repository as

assumed in conventional knowledge management ap-

proaches.

This situation is speciﬁcally accounted for in

the KM framework through encapsulation of well-

differentiated knowledge management functions as

agent roles. Examples proposed in (Langer et al.,

2006) cover amongst others knowledge acquisition,

knowledge processing or brokerage (See Figure 1).

The strength and ﬂexibility of the role abstraction is

that, in contrast to other approaches (Van Elst et al.,

2004), knowledge management-related abilities are

not restricted to highly-specialized dedicated agents.

Rather, any agent is free to assume a time-variant set

of knowledge management roles as deemed appropri-

ate in its situation context. These roles are understood

as auxiliary roles which complement domain-speciﬁc

primary roles such as the management of logistic pro-

cesses. The roles can be further categorized into inter-

nal and external roles. We adopt a notion of internal

roles where these are characterized 1) by reasoning

capabilities and 2) a deliberation pattern (i.e., a plan

to accomplish the knowledge management task). In-

ternal roles can be conducted self-sufﬁciently. Exter-

nal roles by contrast require interaction, structured by

one or more interaction protocols.

3.2 Knowledge Management Roles for

Interactive Adaption

Two complementary roles are involved in interactive

adaption, namely the advisee role played by an agent

seeking assistance in adaption and the advisor role

played by a temporary domain expert. In our mul-

tiagent environment, these roles are no longer tied to

speciﬁc types of actors. We now introduce additional

knowledge management roles and role aspects (See

Figure 1) which extend our framework for distributed

knowledge management (Langer et al., 2006).

The model acquisition role is a specialization of

the knowledge processing role, it presupposes access

TOWARDS MULTIAGENT INTERACTIVE ADAPTION OF INDIVIDUAL MODELS FOR DECISION SUPPORT

555

Knowledge

Consumer

Knowledge

Broker

Knowledge

Provider

Knowledge

Processing

Model

Acquisition

Knowledge

Acquisition

Advice

Integration

Advisor

Advisee

Advisory

Broker

Service

Request Advisory

Service

Request advice for

learning problem

Role Specialization

Role Interaction

Task Delegation

Role Extensions

Basic KM Roles

E E

Figure 1: Adopted roles in the distributed KM framework by Langer et al. (Langer et al., 2006) (gray) and extensions for

interactive adaption of individual models.

...

Advisee

Advisor

Advisee

Advisor

...

Advisee

Advisor

...

Advice

(P)

P,AdvList

Advice

(P)

Advice

(P)

Advice

(P)

Advice

(P)

Advice

(P)

1:1:(n-1)-Interaction

1:n-Interaction

1:1-Interaction

Figure 2: Overview of possible interaction patterns for knowledge transfer episodes.

to representative training data and an adequate ma-

chine learning scheme (e.g., a decision tree learner).

Contingent on the agents’ primary domain role(s), the

data used for machine learning may constitute indi-

vidual experience gathered through action in the task

domain. It may as well originate from a data reposi-

tory accessible to the learner. Once a model such as

a classiﬁer has been learned successfully, we assume

that the role also exposes its inferential capabilities

for internal used by the agent.

The advisee role is a specialization of the knowl-

edge consumer role, elevated to the level of knowl-

edge rather than information. Any agent may as-

sume the advisee role when an assessment of its de-

cision support model has shown deﬁciencies in the

model performance that cannot be handled by inter-

nal means alone. Such a situation may arise due to

an insufﬁcient access to representative training data

such that the prerequisites for the model acquisition

role cannot be met. If the agent is keen to improve

its model immediately rather than wait for new data

to become accessible, the advisee role offers a viable

action alternative. The agent then becomes an active

learner in that it actively seeks for and eventually ap-

proaches peers that assume the learning advisory role

introduced below. In the interaction associated with

these roles, advisors are presented with learning prob-

lems and asked to offer advice to address said problem

based on their models.

The advice integration role is a specialization of

the knowledge processing role. Advice integration is

understood as a subsidiary task to succeed with the

advisee role. We propose that the superordinate ad-

visee role delegates advice integration to this internal

role. One can conceive different feasible interpreta-

tions of this role. As a ﬁrst option, the advice pro-

vided as input may be used to directly revise an ex-

isting model (e.g., by pruning or expanding branches

ICAART 2012 - International Conference on Agents and Artificial Intelligence

556

in a decision tree, or revision of a rule set). A sec-

ond option is to conceive the advice integration as a

specialization of the model acquisition role presented

above. In such a case, a new model is learned based

on the initial training data and the accumulated advice

as additional background knowledge to bias/focus the

operation of a learning scheme that is able to handle

the additional input.

The advisor role is the role complement to the

advisee role and a specialization of the knowledge

provider role. It may be played by any agent with ac-

cess to a particular decision support model if it wants

to provide an knowledge advisory service for agents

which need to learn and maintain similar models. The

advisor model constitutes a valuable asset in that it

captures relevant domain expertise for agents in sim-

ilar application contexts. The model which is used

as basis for the advisory service may be hand-crafted,

yet the probably more interesting use case involves

individually learned models. In order to play an ad-

visor role, it is necessary to interpret requests placed

by advisees and compile tailored advice to address the

communicated learning subproblem, based on the ad-

visor’s own model. The role abstracts from the par-

ticular type of decision support model used by an ad-

visor. While the adoption of an advisee role is an im-

mediate consequence of a concrete need, the comple-

mentary advisor role may be played persistently.

The advisory broker role is an external role. It

is devised as a specialization of the knowledge bro-

ker role. An agent in this role acts as a specialized

yellow pages service within the MAS. It administers

meta-information about knowledge advisory services

exposed by agents currently assuming one or more ad-

visor roles for speciﬁc learning problems. The meta-

information is deposited by the advisors. It speciﬁes

amongst others the respectively handled learning task

(e.g., classiﬁcation), any bias towards particular learn-

ing objectives (e.g., avoidance of false positive classi-

ﬁcations) and a meta-description of the learning do-

main. An advisory broker also accepts requests by

agents seeking advice for a learning problem. The

broker matches the request information against its ad-

visory portfolio in order to point the requesting agents

to suitable interaction partners.

3.3 Structuring Agent Interaction for

Knowledge Transfer Episodes

Further building blocks for our approach are means to

organize interactive adaption. To that end, we propose

a two-layer approach: On the lower layer, an interac-

tion protocol is used to structure the course of a single

knowledge transfer episode. Each such episode spans

the acquisition of a single piece of advice relating to

a speciﬁc learning subproblem. On the upper level,

agents need to structure their global interactive learn-

ing process which may involve multiple consecutive

lower-level transfer episodes. Each is thereby char-

acterized by a speciﬁc learning subproblem which is

brought up by an advisee. In the episode, the advisee

engages in direct interaction with one or more suitable

advisors which need to be known beforehand. Given

such a list of advisors, the advisee can then utilize

several interaction patterns as sketched in Figure 2.

Single-Tier 1:1 Interaction: As a ﬁrst option, the

advisee may choose one advisor for exclusive inter-

action. As a result, the advisee will receive a single

piece of advice to resolve the learning subproblem ad-

dressed in the transfer episode. If the advisor refuses

to advise on the speciﬁed problem, the outcome may

also be a failure of the learning episode.

Single-Tier 1:n Interaction: In order to broaden

advice acquisition for the learning subproblem at

hand, and at the same time reduce to the failure po-

tential of the transfer episode, the agent may place the

same advisory request with a larger number of advi-

sors. As a result of this interaction pattern the ad-

visee may be able to receive multiple independently

compiled pieces of advice. As a consequence, it lies

within the responsibility of the advisee to perform ad-

vice integration. In case of complementary advice,

such an operation may be straightforward. However,

advice integration also involves conﬂict resolution for

contradictions. An interaction protocol for this inter-

action pattern is depicted in Figure 3.

Multi-Tier 1:1:n Interaction: The two interaction

patterns discussed so far are single-tier interactions

in which the advisee interacts directly with all advi-

sors involved in the transfer episode. An alternative is

a multi-tier interaction which again involves a group

of advisors. In contrast to the preceding interaction

patterns, the advisors appear as a holon

with a sin-

gle advisor acting as holon head. The advisory holon

is created dynamically upon request by the advisee.

Besides the learning subproblem that constitutes the

topic of the interaction, the advisee also communi-

cates to an initial advisor, acting as holon head, a set

of additional advisors. These then are to be consulted

while preparing a single, consolidated piece of advice.

The initial advisor uses this information about the ad-

ditional advisors to relay the learning problem. It is

also responsible for advice integration. The interac-

tion among the advisors can itself follow different in-

teraction patterns depending on the preferred method

In the categorization of holonic MAS by Fischer et

al. (Fischer et al., 2003), the organizational form of an mod-

erated association is adequate.

TOWARDS MULTIAGENT INTERACTIVE ADAPTION OF INDIVIDUAL MODELS FOR DECISION SUPPORT

557

: Advisee : Advisor

inform-result: cannot_advise

inform-result : can_advise

(Learning Sub Problem P)

inform-done : inform

request: advice_on(P)

request: can_advise_on(P)

1 n

n-m

m-o

: Advisee : Advisor

inform-result: cannot_explain

inform-result: explanation(Adv)

request: explain_advice(Adv)

inform-done : inform

1 n

p ≤ o

inform-result : advise(P)

(Advice Adv)

Figure 3: Protocol for single-tier 1:n interaction in a knowledge transfer episode.

of advice integration.

A Single-Tier 1:n Interaction Protocol We now

present an interaction protocol for the single-tier 1:n

interaction pattern (See Figure 3). The protocol con-

sists of multiple stages. First, the advisor requests

from all potential advisors, whether or not they con-

sider themselves ﬁt to provide advice for the learning

subproblem at hand. The advisee waits for the respec-

tive advisor responses. Based on this feedback, the

advisee can select the subset of advisors from which

it can then request actual advice. In case the advisors

do not only communicate a simple binary decision

with respect to their willingness to advise, but also

include information about their conﬁdence to provide

’good’ advice, the advisee may choose to continue the

conversation with a selection of all capable advisors.

Those which are excluded are consequently informed

that their services are no longer required. The remain-

ing advisors receive requests to advise. The interac-

tion protocol in Figure 3 then envisages that the earlier

consent to advise enforces that each advisor actually

provides individual advice and does not back out at

this stage of the conversation. In case the communi-

cated advice proves to be comprehensible for the ad-

visee, the interaction with the respective advisor ends

here. If, however, the advisee cannot comprehend any

piece of advice directly, the interaction protocol pro-

vides for a continuation of the conversation to (repeat-

edly) request additional information to further under-

standing.

4 RELATED WORK

Interactive adaption of decision support models bears

a cross-sectional character with related work from

several research ﬁelds, namely machine learning (ML)

and multiagent systems (MAS). In Section 2, we in-

troduced related ML-research, most notably argument

based machine learning by Možina et al. (Možina

et al., 2007). While still restricted to adaption of rule-

based classiﬁcation models (AB-CN2), the approach

has been embraced by Napierala and Stefanowski for

the MODLEM rule induction algorithm (Napierala and

Stefanowski, 2010). We consider these approaches

a vantage point to implement the advice integra-

tion role. Ontañón and Plaza propose an approach

to interactively resolve inconsistencies in individu-

ally learned concept descriptions based on computa-

tional argumentation (Ontañón and Plaza, 2010). As

their goal is the consolidation of individual models,

the approach is interesting when advice from mul-

tiple advisors must be consolidated to reach con-

sensus. Costantini and Tocchio propose an ap-

proach for learning by knowledge exchange in log-

ical agents (Costantini and Tocchio, 2005). The au-

thors concentrate on the exchange of agent beliefs and

rules encoding action recipes. They contemplate the

role of trust in exchanging knowledge and propose

strategies for graded operationalization of acquired

knowledge. Jakob et al. implement adaptability in

MAS by means of collaborative logic-based learning,

focussing on communication strategies for acquired

knowledge (Jakob et al., 2008). At the cross-section

ICAART 2012 - International Conference on Agents and Artificial Intelligence

558

of MAS, robotics, and reinforcement learning, the im-

plementation of social learning strategies have drawn

signiﬁcant attention (see, for instance, (Noble and

Franks, 2003)).

5 DISCUSSION AND FUTURE

WORK

In this paper we have sketched an interaction-based

approach to the adaption of individual decision sup-

port models in MAS. It is desirable when software

agents in complex, dynamic environments need to up-

date, adapt, or improve their knowledge base for de-

cision making. Sometimes, this improvement process

can be based on machine learning from observational

data, alone. But when available data is insufﬁcient

in quantity or quality, when data is too expensive, or

when the machine learning process turns out to be too

complex, alternative approaches are needed. There

are two basic components in our approach: 1) a set of

speciﬁc knowledge transfer roles which extends a set

of basic knowledge management roles, and 2) a col-

lection of interaction protocols for knowledge trans-

fer.

Ongoing work on the proposed multiagent frame-

work comprises a prototype implementation of the

KM roles and interaction patterns introduced in Sec-

tion 3. We focus ﬁrst on the adaption of rule-based

classiﬁcation models, relying, for advice integration,

on the ABML approach by Možina et al. In the pro-

cess, we will also elaborate the meta-control, used

by an advisee to guide its interactive adaption pro-

cess over multiple knowledge transfer episodes, as a

ﬂavor of local search in a model space. The proto-

type is implemented based on the JADE agent devel-

opment environment. Evaluation will be performed

in the PlaSMA multiagent-based simulation environ-

ment (Warden et al., 2010). In the future, we also

seek to enable a more far-reaching interoperability be-

tween heterogeneous agents in the context of inter-

active model adaption. This includes support for di-

versity in employed models (e.g., rule-based for the

advisee and ANN-based for the advisor(s)). It also in-

cludes support for heterogeneity in the training data

(with respect to attributes), discretization of values in

individual learning, and the naming of attributes and

concept classes. These extensions speciﬁcally call for

the provision of additional KM roles, enabling, for in-

stance, semantic mediation.

REFERENCES

Costantini, S. and Tocchio, A. (2005). Learning by Knowl-

edge Exchange in Logical Agents. In WOA, pages 1–

8. Pitagora Editrice Bologna.

Fischer, K., Schillo, M., and Siekmann, J. (2003). Holonic

Multiagent Systems: A Foundation for the Organisa-

tion of Multiagent Systems. In Holonic and Multi-

Agent Systems for Manufacturing, pages 1083–1084.

Springer.

Gehrke, J., Herzog, O., Langer, H., Malaka, R., Porzel, R.,

and Warden, T. (2010). An Agent-Based Approach

to Autonomous Logistic Processes. Künstliche Intel-

ligenz, 24(2):137–141.

Jakob, M., Toži

cka, J., and P

echou

cek, M. (2008). Collab-

orative Learning with Logic-Based Models. In Adap-

tive Agents and Multi-Agent Systems III. Adaptation

and Multi-Agent Learning, volume 4865 of LNCS,

pages 102–116. Springer.

Jennex, M. (2009). Re-Visiting the Knowledge Pyramid. In

Proceedings of the 42nd Hawaii Intl. Conference on

System Sciences, 2009., pages 1–7. IEEE.

Langer, H., Gehrke, J. D., Hammer, J., Lorenz, M., Timm,

I., and Herzog, O. (2006). A Framework for Dis-

tributed Knowledge Management in Autonomous Lo-

gistic Processes. Intl. Journal of Knowledge-Based &

Intelligent Engineering Systems, 10(4):277–290.

Možina, M., Zabkar, J., and Bratko, I. (2007). Argu-

ment Based Machine Learning. Artiﬁcial Intelligence,

171(10-15):922–937.

Napierala, K. and Stefanowski, J. (2010). Argument Based

Generalization of MODLEM Rule Induction Algo-

rithm. In Rough Sets and Current Trends in Comput-

ing, volume 6086 of LNCS, pages 138–147. Springer.

Noble, J. and Franks, D. W. (2003). Social Learning in

a Multi-Agent System. Computing and Informatics,

22(6):1001–1015.

Ontañón, S. and Plaza, E. (2010). Concept Convergence in

Empirical Domains. In Proceedings of the 13th Intl.

Conference on Discovery Science, pages 281 – 295.

Pan, S. and Yang, Q. (2009). A Survey on Transfer Learn-

ing. IEEE Transactions on Knowledge and Data En-

gineering, pages 1345–1359.

Settles, B. (1995). Active Learning Literature Survey. Sci-

ence, 10(3):237–304.

Van Elst, L., Dignum, V., and Abecker, A., editors (2004).

Agent-Mediated Knowledge Management: Interna-

tional Symposium (AMKM ’03), Standford, CA, USA.

Springer. Revised and Invited Papers.

Warden, T., Porzel, R., Gehrke, J., Herzog, O., Langer, H.,

and Malaka, R. (2010). Towards Ontology-based Mul-

tiagent Simulations: The PlaSMA Approach. In Proc.

of the 24th European Conference on Modelling and

Simulation, pages 50–56.

TOWARDS MULTIAGENT INTERACTIVE ADAPTION OF INDIVIDUAL MODELS FOR DECISION SUPPORT

559