Towards Descriptive and Prescriptive Double-Loop Learning Agents

The rise of complex computational models is due to the desire for white-box models with higher resolution of explanation and representation. Usually, the reason for complexity within models is because we are trying to explain real world phenomena that include humans. Descriptive qualitative and quantitative models of human behavior have mainly been the goal of social sciences (psychology, economy, sociology etc.) as well as fields such as cognitive science. Most of the time, studies are primarily interested in behavior within a specific context or situation in that domain. Therefore, generated theories are restricted to apply within the domains that they are designed for, constrained by further assumptions. Hence, often a single theory is not sufficient to properly represent human behavior in an evolving or dynamic socio-economic systems model. This makes a systems approach that contains adaptive feedback mechanisms to this problem necessary. A possible framework that highlights this kind of mechanism is double-loop learning (Argyris and Schon, 1978) (Figure 1). The first loop of learning is based on an existing mental model (Johnson-Laird, 1983). A mental model is an implicit internal image of how the current system works (Senge, 1990). In other words, mental models can be interpreted as the theory that results in a strategy or decision making mechanism such as a heuristic. Most of the behavioral theories and heuristics can be interpreted as possible mental models that we utilize under certain conditions. The single loop learning only considers the existing mental model and modifies it based on information fed back, i.e. consequences of our actions. In this loop, the way we view the world does not change and we just make fine tuning adjustments on the existing mental model. The second loop of learning is where we consider whether our current mental model is still satisfactory to explain the world dynamics or not. Within a system, certain behavior around us might lead to a paradigm shift in our explanation or we might explain certain situations with one mental model and other situations with other sets of mental models. Hence, our reasoning mechanism adapts to the changes in the world. Although humans are capable of doing doubleloop learning, none of these learning loops is done perfectly. Therefore, a descriptive human behavior model based on a double-loop learning framework would have to reflect human faults in application. A prescriptive approach would point to our faults in the learning processes and in our mental models. Hence, an ideal agent would utilize correct mental models, rules, and/or heuristics at the right time with correct settings in a complex system. In this study, I plan to provide a general framework for descriptive and prescriptive (ideal) models of double-loop learning.