Evidentialist Foundationalist Argumentation in Multi-Agent Systems

This dissertation focuses on the explicit grounding of reasoning in evidence directly sensed from the physical world. Based on evidence from human problem solving and successes, this is a straightforward basis for reasoning: to solve problems in the physical world, the information required for solving them must also come from the physical world. What is less straightforward is how to structure the path from evidence to conclusions. Many approaches have been applied to evidence-based reasoning, including probabilistic graphical models and Dempster-Shafer theory. However, with some exceptions, these traditional approaches are often employed to establish confidence in a single binary conclusion, like whether or not there is a blizzard, rather than developing complex groups of scalar conclusions, like where a blizzard's center is, what area it covers, how strong it is, and what components it has. To form conclusions of the latter kind, we employ and further develop the approach of Computational Argumentation. Specifically, this dissertation develops a novel approach to evidence-based argumentation called Evidentialist Foundationalist Argumentation (EFA). The method is a formal instantiation of the well-established Argumentation Service Platform with Integrated Components (ASPIC) framework. There are two primary approaches to Computational Argumentation. One approach is structured argumentation where arguments are structured with premises, inference rules, conclusions, and arguments based on the conclusions of other arguments, creating a tree-like structure. The other approach is abstract argumentation where arguments interact at a higher level through an attack relation. ASPIC unifies the two approaches. EFA instantiates ASPIC specifically for the purpose of reasoning about physical evidence in the form of sensor data. By restricting ASPIC specifically to sensor data, special philosophical and computational advantages are gained. Specifically, all premises in the system (evidence) can be treated as firmly grounded axioms and all arguments' conclusions can be numerically calculated directly from their premises. EFA could be used as the basis for well-justified, transparent reasoning in many domains including engineering, law, business, medicine, politics, and education. To test its utility as a basis for Computational Argumentation, we apply EFA to a Multi-Agent System working in the problem domain of Sensor Webs on the specific problem of Decentralized Sensor Fusion. In the Multi-Agent Decentralized Sensor Fusion problem, groups of individual agents are assigned to sensor stations that are distributed across a geographical area, forming a Sensor Web. The goal of the system is to strategically share sensor readings between agents to increase the accuracy of each individual agent's model of the geophysical sensing situation. For example, if there is a severe storm, a goal may be for each agent to have an accurate model of the storm's heading, severity, and focal points of activity. Also, since the agents are controlling a Sensor Web, another goal is to use communication judiciously so as to use power efficiently. To meet these goals, we design a Multi-Agent System called Investigative Argumentation-based Negotiating Agents (IANA). Agents in IANA use EFA as the basis for establishing arguments to model geophysical situations. Upon gathering evidence in the form of sensor readings, the agents form evidence-based arguments using EFA. The agents systematically compare the conclusions of their arguments to other agents. If the agents sufficiently agree on the geophysical situation, they end communication. If they disagree, then they share the evidence for their conclusions, consuming communication resources with the goal of increasing accuracy. They execute this interaction using a Share on Disagreement (SoD) protocol. IANA is evaluated against two other Multi-Agent System approaches on the basis of accuracy and communication costs, using historical real-world weather data. The first approach is all-to-all communication, called the Complete Data Sharing (CDS) approach. In this system, agents share all observations, maximizing accuracy but at a high communication cost. The second approach is based on Kalman Filtering of conclusions and is called the Conclusion Negotiation Only (CNO) approach. In this system, agents do not share any observations, and instead try to infer the geophysical state based only on each other's conclusions. This approach saves communication costs but sacrifices accuracy. The results of these experiments have been statistically analyzed using omega-squared effect sizes produced by ANOVA with p-values < 0.05. The IANA system was found to outperform the CDS system for message cost with high effect sizes. The CDS system outperformed the IANA system for accuracy with only small effect sizes. The IANA system was found to outperform the CNO system for accuracy with mostly high and medium effect sizes. The CNO system outperformed the IANA system for message costs with only small effect sizes. Given these results, the IANA system is preferable for most of the testing scenarios for the problem solved in this dissertation