Complex Task Solving in Groups of Autonomous and Collaborative Agents...

We model a system of autonomous and collaborative decision-making agents who self-organize themselves in groups in order to solve complex tasks. Previous literature argues that such groups should be static, in terms of not periodically adapting to the environment by replacing its members, since it is detrimental to the groups’ performances [1]. In contrast, other lines of research suggest that static group structures do not necessarily bring a better performance than dynamic group structures [2]. As these findings look contradictory, our objective is to better understand the effects that static and dynamic group structures have on group performance. 

In order to do so we incorporate the idea of adaptation [3] into the autonomous formation of groups. We set up an agent-based model and base the task environment on the NK-framework [4], which we adapt for complex task solving in groups (see also [5]). Dynamic (and static) features of groups are studied at two levels: First, we consider various degrees of individual adaptation to the task environment by endowing agents with capabilities to learn new partial solutions to the complex decision-making problem. Second, by allowing for the autonomous reorganization of group structures using an auction-based mechanism, we consider different extents of collective adaptation at the level of the group. 

The results of our simulation study suggest that there are complex trade-offs between individual and collective adaptation and task performance. For scenarios with low task complexity, we find that a high level of individual adaptation improves performance significantly, whereas collective adaptation has only marginal effects. For highly complex tasks, however, we can observe that whether a high frequency of collective adaptation is desirable or not depends on the extent of individual adaptation. If individual adaptation is low, collective adaptation results in better group performance. Collective adaptation, however, can be detrimental to the performance of groups formed by agents who are already endowed with well-performing learning mechanisms. Based on our results, we provide advice for the efficient design of systems composed of autonomous and collaborative agents. 

References 

[1] S.C. Hsu, K.W. Weng, Q. Cui, W. Rand, Understanding the Complexity of Project Team Member Selection Through Agent-Based Modeling, International Journal of Project Management, 34:1 (2016), 82-93. 

[2] L. Sless, N. Hazon, S. Kraus, M. Wooldridge, Forming K Coalitions and Facilitating Relationships in Social Networks. Artificial Intelligence, 259 (2018), 217-245. 

[3] D.J. Teece, G. Pisano, A. Shuen, Dynamic Capabilities and Strategic Management: Organizing for Innovation and Growth. Strategic Management Journal, 18:7 (1997), 509-533. 

[4] D.A. Levinthal, Adaptation on Rugged Landscapes. Management Science, 43:7 (1997), 934-950. 

[5] D. Blanco, S. Leitner, A. Rausch, Dynamic Coalitions in Complex Task Environments: To Change or Not to Change a Winning Team?, preprint from https://arxiv.org/abs/2010.03371 (2020).