Epistemic Structure: An Inquiry into How Agents Change the World for Cognitive Congeniality. Technical Report 2005-02

This thesis presents cognitive mechanisms that explain how humans and other organisms generate epistemic structures (ES) in the environment. Epistemic structures are structures generated systematically in the world by organisms to minimise cognitive load, for oneself, others, or both. Examples of ES in non-human organisms include pheromones, markers etc. For humans they include labels, colour codes etc. Adding structures to the world for cognition is a fundamental adaptive strategy that exists across species. So a basic mechanism, growing in complexity, is required to explain how the strategy works in different species, from ants to humans. Such a mechanism is proposed, starting from low-level organisms and building up to humans. The model for lower organisms proposes that they learn the epistemic (i.e. knowledge) value of structures they inadvertently generate in the world (like pheromones). This learning of epistemic value is based on a feedback of cognitive load. Results from a proof-of-concept multi-agent simulation, based on the Q-learning algorithm, provide support for the model. This model is then extended to the human case, using a set of theoretical arguments and examples, accounting for situations where humans generate ES for themselves and others. The case where structures are generated exclusively for others requires a more complex model. Based on further arguments and evidence, mental simulation of action is proposed as the mechanism underlying this case. This hypothesis is tested using a scenario-based methodology, adapted from counterfactual thinking research. Participants are asked to provide solutions to real-life problems involving a series of actors, including iii robots and dementia patients. Results indicate that simulation is the stronger candidate mechanism underlying ES generation just for others. The final section examines the robustness of the epistemic structure strategy, using the passing problem in the robotic soccer simulation environment. The experiments show that adding ES to the world (yells) increases passing accuracy by 8-17 percentage points, compared to a centralised decision-making strategy. The ES strategy is also shown to perform better in high-noise and high-processing load situations, indicating that this robustness could act as a driving mechanism in the evolution of ES as a survival strategy