A Multi-Agent Spreading Activation Network Model

for Online Learning Objects

Ashraf Saad, PhD

Computer Engineering

Georgia Institute of Technology

GTREP, Savannah, GA 31405

Abstract

The development of an online learning environment depends on the underlying knowledge representation. Learning Objects have been promoted by various leading organizations, such as IEEE, as a viable building block for the knowledge representation of educational resources for learning environments in general. However, one of the most important educational questions pertaining to the use of Learning Objects that is still unanswered is: how will such objects be interconnected, to form more complex knowledge modules, in order to achieve certain desired learning outcomes for a particular learner? The work described in this paper presents an approach for interconnecting Learning Objects into learner-specific Learning Paths via a Spreading Activation mechanism. The resulting Network model for interconnecting Learning Objects therefore supports Multi-Agent Spreading Activation that is applicable to online learning environments. In doing so, this work aims to accomplish the personalization of knowledge acquisition for each learner in a Learning Objects-based learning environment.

Keywords: learning objects, spreading activation, network model, decision-theoretic.

  1. Learning Objects

The concept of Learning Objects [3] has been promoted by several leading and standards-setting organizations as a viable approach to standardize the creation of educational content libraries and learning management systems. Such organizations and their related activities include:

1)The IEEE (Institute of Electronics and Electrical Engineers) Learning Technology Standards Committee (LTSC): P1484 set of Working Groups and Study Groups and the resulting working documents, to-date [7].

2)The IMS (Instructional Management System) Global Learning Consortium: developing and promoting open specifications for facilitating online distributed learning activities [8].

3)The US Federal Government: Advanced Distributed Learning (ADL) initiative and the resulting Shareable Courseware Object Reference Model (SCORM) [9].

4)The European Union: PROMETEUS (PROmoting Multimedia access to Education and Training in EUropean Society) programme [10] and the ARIADNE (Alliance of Remote Instructional Authoring and Distribution Networks for Europe) project [11].

As a result of these projects and initiatives, a standardized approach for specifying Learning Objects and their associated metadata is bound to emerge.

While the standardization of educational content provided by Learning Objects is an important goal, a more vital goal for learning environments is the personalization of the learning process to the needs of a particular learner [6]. This work aims to accomplish the personalization of knowledge acquisition for each learner in a Learning Objects-based online learning environment.

  1. Spreading Activation

Spreading Activation is a well-studied technique by cognitive scientists seeking to understand the learning processes that take place to form learning networks. The origin of these studies lies in the study of how [biological] learning networks are actually formed. The origins of spreading activation can be traced back to early work on associative memory at the end of the nineteenth century. Researchers in the field of Artificial Intelligence have emulated this process in a variety of ways and applied spreading activation over a semantic network to solve several important problems in Artificial Intelligence [2]. Such an approach has also been applied for machine learning in engineered systems, such as in the field of robotics [1].

Some of the main characteristics of Spreading activation include that are important to this work are [2]:

  1. Spreading activation subsumes both marker passing and local connectionism.
  2. Since spreading activation is a distributed process, information must be localized at each node in the network.
  3. Spreading activation is therefore the passing of messages – of arbitrary complexity – between concurrent objects – of arbitrary complexity.

Recently, spreading activation has been applied to robotic task planning under uncertainty [1]. This novel approach to task planning utilizes a decision-theoretic framework where action selection is driven by a spreading activation mechanism on a probabilistic network that encodes the domain knowledge. The main characteristics of this approach that are important to this work are [1]:

  1. It utilizes a semantic network that comprises both proposition and action nodes that are interconnected via probabilistic links; i.e. the weight associated with each defines the correlation between the success of an action and its preconditions.
  2. Spreading activation for action selection utilizes both forward propagation from the current state as well as backward propagation of the goal utility.
  3. Since spreading activation is done over a probabilistic network, action selection therefore leads to decision-theoretic planning: the utility received by an action represents the product of the probabilities of success of the subsequent actions in the path leading to the goal proposition.
  4. The trade off between the reliability of an action versus its cost is also explicitly addressed by incorporating the cost (expressed in terms of any resource of interest) of the action as well as its probability of success.
  1. Spreading Activation over a Network Model of Learning Objects

In the context Learning Objects, the necessary metadata will need to be associated with each object in order to enable a Spreading Activation mechanism to automatically generate Learning Paths that are tailored to a specific learner, and leading to realizing the learner’s particular learning objectives. A semantic network of learning objects can be formed in order to utilize a spreading activation mechanism to generate the desired learning paths, as illustrated in Figure 1.

LO = Learning Object node

C = Proposition node (Precondition)

In the section of the network shown in Figure 1, the directional weighted links connect:

  1. The output from a proposition node to the input of a learning object.
  2. The output from a learning object to the input of a proposition node.

Some of the main characteristics of the proposed network model include the following ones:

  1. Each learning object corresponds to a basic “quantum” of knowledge. The entire

Network, both learning objects and propositions, represents the subject knowledge

that a learner needs to acquire.

  1. A learning object can receive input from one or more proposition nodes and can send

an output to one or more proposition nodes.

  1. A proposition node can receive input from one or more learning objects and can send

an output to one or more learning objects.

  1. A proposition node is both the effect of the learning object(s) preceding it, and a

precondition for the learning object(s) following it.

  1. Forward propagation proceeds from the knowledge state of the learner at any given

point in time.

  1. Backward propagation proceeds from the desired state of knowledge for the learner.
  2. Spreading activation persists until the knowledge state of the learner equals the

desired knowledge state as represented in the network.

  1. The weights of the links between learning objects and propositions can be learned

and adaptively fine tuned based on the success of the network in fulfilling the desired

learning objectives for a variety of learners.

  1. Multi-Agent Learning Environments

The proposed network model for knowledge representation lends itself to a Multi-Agent Learning Environment. It can support, and can be used at the basis, of the development of the following agents:

  1. An Agent-based Student Modeler that captures the cognitive state, as well as the learning preference(s)/style(s), of the learner and tracks them over time [4], identifying any deficiencies or misconceptions in learning the subject knowledge.
  2. An Agent-based Intelligent Tutoring System [5] that can utilize the output of the Student Modeler to guide and aid the learner through the learning process, and that attempts to correct the learners deficiencies or misconceptions in learning the subject knowledge.
  3. A Curriculum Generation and Management Agent [5] that can generate lesson plans for a variety of learners.
  4. A Social Agent that can enable learners to collaborate on solving problems.

Knowledge Query Manipulation Language (KQML) [12] can be used as the basis for communications between these agents.

  1. Conclusions and Directions for Future Work

Some of the important questions that still need to be answered for the proposed approach include:

-What is the metadata necessary for a Learning Object to be comprised in a semantic network that utilizes the proposed spreading activation mechanism.

-How will the metadata account for pedagogical; e.g. learner style and instructional approach, versus technological; e.g. wired/wireless connectivity and the associated bandwidths and hardware/software capabilities, issues?

-What are the limitations of applying Spreading Activation to a Multi-Agent architecture?

Work in the immediate future will focus on applying spreading activation to a learning objects-based semantic network for a specific subject knowledge from the computer engineering discipline. Two of the most viable technologies available today to implement the necessary computational framework include XML [13] and JatLite [14].

References

[1] S. Bagchi, G. Biswas, and K. Kawamura, “Task Planning under Uncertainty using a Spreading Activation Network”, IEEE Transactions on Systems, Man and Cybernetics, Vol. 30, No. 6, November 2000, pp 639-650.

[2] J-P. Corriveau, “On the Design of a Concurrent Object-Oriented Spreading Activation Architecture”, Proceedings of the Twenty-Seventh Annual Hawaii International Conference on System Sciences, January 1994, pp 73-81.

[3] M. Martinez, “Designing Learning Objects to Mass Customize and Personalize Learning.” In Wiley, D. (ed.), Instructional Use of Learning Objects. Association for Educational Communications & Technology, November, 2000.

[4] G. McCalla, J. Vassileva, J. Greer, and S. Bull, “Active Learner Modeling”, Fifth International Conference on Intelligent Tutoring Systems, Montreal, Quebec, Canada, June 19-23, 2000.

[5] N. Capuano, M. Marsella, and S. Salerno, “ABITS: An Agent Based Intelligent Tutoring System for Distance Learning”, Fifth International Conference on Intelligent Tutoring Systems, Montreal, Quebec, Canada, June 19-23, 2000.

[6] M. Martinez and C.Victor Bunderson, “Foundations for Personalized Web Learning Environments”, ALN Magazine Volume 4, Issue 2, December 2000.

[7] IEEE Learning Technology Standards Committee (LTSC), P1484 Working Groups and Study Groups, “Standard for Information Technology --Education and Training Systems -- Learning Objects and Metadata”, 2000.

[8] Instructional Management System (IMS) Global Learning Consortium, “IMS Learning Resource Meta-data Best Practices and Implementation Guide, Version 1.0 - Final Specification”, 2001.

[9] R. Cover, “Shareable Courseware Object Reference Model Initiative (SCORM).” In R. Cover, Managing Editor, The XML Cover Pages, 2001.

[10] PROmoting Multimedia access to Education and Training in EUropean Society (PROMETEUS) programme, 2000.

[11] Alliance of Remote Instructional Authoring and Distribution Networks for Europe (ARIADNE) project, 2000.

[12] T. Finin, Y. Labrou, and James Mayfield, “KQML as an agent communication language”, in Jeff Bradshaw (Ed.), ``Software Agents'', MIT Press, Cambridge, 1997.

[13] XML Specifications, 2001.

[14] JatLite, Stanford University, 2001.