Learning About Learning: Using Multi-Agent Computer Simulation to Investigate Human Cognition
Last modified: May 5, 2006
Understanding “how people learn” has been the cornerstone of a significant part of the research in cognitive sciences and education. Understanding the “how” question is a prerequisite to design learning environments in which students learn with higher intellectual and emotional satisfaction. However, there are overwhelming experimental obstacles that cause a definite answer to be still far in the future. On one hand, tools such as fMRIs cannot yet offer the speed and resolution needed to evaluate any complex learning process. On the qualitative side, ethnographic or micro-genetic methods are far from offering a solid, generalizable, task-independent account on how students learn.
The present research agenda proposes a novel methodology using multi-agent simulation (MAS) to explore and investigate human cognitive development and learning. It does not intend to replace current methods, but rather be a bridge between quantitative and qualitative descriptions. The ultimate goal is to enable researchers to generalize and play “what-if” scenarios departing from in-depth interviews and ethnographic data, as well as investigate internal cognitive structures departing from external, observed behaviors. Our work builds on previous seminal contributions to field, in which theoretical models of cognition were implemented in the form of computer programs in attempt to predict human reasoning (Newell & Simon, 1972; Rose & Fischer, 1999). In particular, multi-agent simulation (MAS; e.g., ‘NetLogo,’ Wilensky, 1999) could offer powerful methods for exploring the emergence of self-organized hierarchical organization in human cognition, enabling theoreticians to assign rules of behavior to computer “agents,” whereupon these entities act independently but with awareness to local contingencies, such as the behaviors of other agents.
This paper describes one of the strands in this research effort: we built multi-agent computer models to simulate different cognitive and epistemological modes by which student function in a typical classroom scenario. The models have two types of agents: retrievers and connectors. The first type retrieves information from different sources (books, teacher, previous knowledge, etc.). The second agent, the connector, tries to make sense of the information by evaluating the fit of possible links accordingly to a probabilistic model.
Our working hypothesis, which is being confirmed by our preliminary results so far, is that traditional schooling favors “weak” and fast connectors, i.e., quickly putting together answers for questions without deep reflection or use of previous student knowledge. This strategy, as our computer experiments suggest, is effective for simple content. For complex content with a high number of “pieces”, the number of possible connection grows exponentially. Thus, for more elaborated content, our results suggest that it “pays off” to invest on teaching connecting skills, instead of information retrieving skills. We are currently developing methodologies to juxtapose the results from the models with data collected in classrooms.