Multiagent Coalition Formation for Computer-Supported Cooperative Learning

Leen-Kiat Soh, Nobel Khandaker, Hong JiangAAAI 2006

Abstract

This paper presents I-MINDS, an intelligent multiagent system designed to enhance collaborative learning in large-scale and distance-learning environments. By leveraging a novel Vickrey auction-based algorithm called VALCAM, I-MINDS automates the formation of student "coalitions" (groups) that evolve based on performance and peer compatibility. Experimental results from real-world deployments demonstrate that students using I-MINDS can outperform those in traditional face-to-face settings, with the system effectively leveling the playing field across diverse student cohorts.

Introduction

Collaborative learning, particularly the Jigsaw model, is a proven pedagogical strategy where students become "experts" in sub-topics to teach their peers. However, managing these interactions in large or remote classrooms is a significant challenge for instructors. Traditional Computer-Supported Cooperative Learning (CSCL) systems often act as passive communication channels, failing to assist in group management or partner selection.

I-MINDS addresses this gap by serving as an active participant. It automates group formation, prioritizes student questions, and profiles behavior to ensure effective collaboration, even in the absence of physical social cues.

System Architecture

The I-MINDS infrastructure is built upon three specialized agent types that work in concert to manage the classroom environment:

  • Teacher Agents: Assist instructors by ranking and grouping student questions using Natural Language Processing (NLP) tools. This allows teachers to address the most critical or common inquiries efficiently.
  • Student Agents: Act as personal digital assistants for each learner. They maintain profiles of student interests and performance and participate in automated auctions to secure group placements.
  • Group Agents: Monitor collaborative sessions in real-time, tracking metrics such as message frequency and peer evaluations to assess team health.
Figure 1: The hierarchical agent architecture of the I-MINDS system, showing the relationships between Manager, Teacher, Student, and Group Agents.

The VALCAM Algorithm

The core of the I-MINDS system is the VALCAM (Vickrey Auction-Based Learning-Enabled Coalition and Adaptation for Multiagent Systems) algorithm. VALCAM facilitates group formation through a market-based approach:

  1. Virtual Currency: Students earn currency based on individual and collective performance.
  2. Bidding: Student Agents use this currency to bid on joining specific coalitions.
  3. Vickrey Mechanism: By using a second-price auction, the system encourages truthful bidding, as the winner pays the price of the second-highest bid.
  4. Learning and Adaptation: Over time, agents learn peer compatibility and competence, aiming for "good-enough" groups quickly rather than searching for computationally expensive optimal solutions.

Experimental Evaluation

The system was evaluated through a control-treatment protocol at the University of Nebraska-Lincoln during the Spring and Fall 2005 semesters. Participants were students in an introductory computer science course (CS1). The control group conducted Jigsaw activities face-to-face, while the treatment group used I-MINDS exclusively for all communications and collaborative tasks.

Results and Discussion

Empirical data suggests that the digital environment provided by I-MINDS may actually enhance learning through the "articulation effect." By requiring students to type their thoughts into forums or digital whiteboards, they were forced to clarify their ideas more explicitly than in spontaneous speech.

Figure 6: Comparison of academic performance (Mean and Standard Deviation) between Control and Treatment sections across Spring and Fall 2005 sessions.

Key findings include:

  • Academic Performance: In Spring 2005, the I-MINDS group achieved a mean post-test score of 7.57, significantly higher than the control group's 6.96.
  • Reduced Variance: The standard deviation in the treatment group was consistently lower (e.g., 1.11 vs. 1.25 in Fall 2005), suggesting the system helped "level the playing field" for all students.
  • Peer Compatibility: As the VALCAM algorithm gathered more data, peer ratings within groups improved, indicating more successful pairings over time.

Interestingly, students in the I-MINDS group reported lower "self-efficacy" despite higher actual test scores. This suggests that the ease of face-to-face interaction can sometimes mask a lack of actual academic progress, whereas the digital system provides a more rigorous learning environment.

Conclusion and Future Work

While I-MINDS showed strong results, limitations remain regarding the user interface and the lack of non-verbal cues. Future work focuses on integrating I-MINDS directly into Integrated Development Environments (IDEs) for real-time coding collaboration and refining the VALCAM algorithm to better balance subjective peer reviews with objective performance data.

Summary updated

BibTeX
@inproceedings{soh2006multiagent,
  author = {Leen-Kiat Soh and Nobel Khandaker and Hong Jiang},
  title = {{Multiagent Coalition Formation for Computer-Supported Cooperative Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure