1. GC-MAS - a Multiagent System for Building Creative Groups used in Computer Supported Collaborative Learning Gabriela Moise, Monica Vladoiu, Zoran Constantinescu

2. Subject method for building creative teams, based on unsupervised learning and with support from a multiagent system first experiments on grouping learners involved in online brainstorming 2

3. Computer Supported Collaborative Learning (CSCL) has appeared as a reaction to software used previously in learning, which have been forcing learners to study and learn as isolated individuals in CSCL, learning is obtained by computer- supported interactions both between learners and between learners and teachers 3

4. GC-MAS - A Multiagent System for Building Creative Groups 4

5. The Communication Agent (CommGC): interfacing with the users and with the agents managing the activities of the other agents The Creative Groups’ Builder (BuildGC): construction of the creative groups (unsupervised learning algorithm, classification techniques) The Creativity Evaluation Agent (EvalGC): support for assessment of group creativity The Creativity Booster (EnvrGC): stimulates creative contextual environments that provide for increasing group creativity 5

6. The Glue Role Agent (GlueGC): seeking out and taking on otherwise neglected tasks that have potential to facilitate creative group performances The Facilitator Agent (FCL-GC): supports the facilitator in helping groups to interact more efficiently The Team Relational Support (TRS-GC): supports the team members in providing support for the other group members 6

7. BuildGC construction and iterative refinement of creative groups taking into account the components that generate creativity their interdependencies that have effect on creativity the purpose of building of creative groups (to solve a problem, to complete a task etc.) The current reasoning process is based on an adapted version of the Q-learning algorithm 7

8. In brief, this algorithm is a reward learning algorithm that starts with an initial estimate Q(s, a) for each pair. When a certain action a is chosen in a state s, the system (BuildCG) gets a reward R(s, a) and the next state of the system is acknowledged 8

9. in our case, we tackle n students for each student, a characteristic vector that includes m individual features is constructed, namely (c 1, c 2, …, c m ) a state consists of this vector and the group number, while an action refers to moving a student to another group Q expresses the quality of association between a state and an action our goal is to build the most creative k groups (k being given) 9

10. GC-Q-learning adapted algorithm build a bi-dimensional matrix Q for all the possible pairs. The columns of this matrix consists of (c 1, c 2, …, c m, no_group, action_number, q) initialize the optim_policy (in our case is the optimal grouping) with a guided policy, and Q_optimal with Q 10

11. GC-Q-learning adapted algorithm group the students and undertake working sessions (in our first experiments, online brainstorming), in which the group creativity is assessed and its score is assigned to R(s,a) – using EvalGC. For each such working session, the matrix Q is calculated analyze matrix Q. The optimal policy is given by the action for which Q_optimal gets the maximum value 11

12. GC-Q-learning adapted algorithm once the optimal policy consisting in tuples (c 1, c 2, …, c m, group number) is obtained, predictions for each set of data can be made based on advanced classification techniques (Bayesian networks, neural networks etc.) the Q values are the same for all the members of a group 12

13. Sample data Our goal: to group in increasingly creative groups 12 students having the (Gough, motivation) as follows: 3 students with (3,1) 4 students with (3,2) 2 students with (2,1) 1 students with (1,1) 2 students with (4,1) 13

14. 14 GMGMGMGM Group 131323131 Group 221222132 Group 341413232 Group 432213232 Group 531213241 Group 631312241 Group 731312141 Group 832213232 Group 931224132

15. Sample data 15 NicknameGough scoreMotivationGroup_noReward Spiderman3113 Curly-eirene3213 Nightyzor3113 Ash3113 Trol2124 Heathcliff2224 LiGiK2124 midgey3224 Peaches4134.33 HARA FTW!4134.33 Ashley3234.33

16. Sample data - interpretation A student with (3,1) would be the most creative if s/he would be in group 7, and decreasing - in group 1, 6, 5 or 9 16 GoughMotivation Action- group-noQ 3113.46875 3120 3130 3140 3152.697188 3163.295781 3173.798281 3180 3192.532188

17. Sample data - interpretation A student with (3,2) would be the most creative if s/he would be in group 8, and decreasing - in group 4, 3, 9, 2 or 1 17 GoughMotivation Action- group-noQ 3211.5 3222.375 3234.234063 3244.477402 3252.888516 3260 3270 3284.872614 3292.883153

18. Conclusions and future work we introduced here our semi-automated method of grouping team members in increasingly creative groups future work: corroborating the results obtained with several creativity evaluation scales, assessment of creativity before and after activities assumed to help trigger creativity, inclusion of contextual and organizational factors, testing the method in other activities, improving of the algorithm, offering the method as an online open service 18

19. Thank you! 19