Abduction, or inference to the best explanation, is, plausibly, part of commonsense reasoning, and a means by which a cognitive system may arrive at estimates of its world from observational and other evidence. We take this "world estimate" to be the cognitive system's beliefs. Since such reasoning is fallible, and world estimates will sometimes contain errors, an abductive reasoning system might improve its performance if it has a way to engage in belief revision when new evidence, or further reasoning, indicates the existence of a problem.
In this study, we develop, implement, and experimentally validate a metareasoning system that monitors and attempts to correct beliefs established by the base-level abductive reasoning system. We first identify that the presence of an anomaly, which we define as an observation or other evidence that cannot plausibly and consistently be explained, as a signal that the cognitive system's world estimate might be incorrect or, alternatively, that the unexplainable datum is noise. The metareasoning system responds to the presence of anomalies by asking exactly that question: which anomalies are due to mistakes in the world estimate, and warrant specific belief revisions, and which anomalies are due to noise, and should not instigate belief revisions? Various considerations regarding the nature of the anomalies and the system's reasoning history are brought to bear to answer this question.
Fundamentally, we see the metareasoning question ("what explains these anomalies: mistaken beliefs, or noise?") as structurally similar to the cognitive system's original question, "what explains these observations?" Thus, the metareasoning system is an abductive reasoning system, just like the base-level system. The anomalies constitute meta-evidence which may be explained by meta-hypotheses. These meta-hypotheses describe the various kinds of causes of anomalies and specify particular belief revisions in order to resolve the anomalies. The same abductive reasoning algorithms employed by the base-level reasoner are activated to find the best explanation for the anomalies. An anomaly is judged to be the result of noise when no meta-hypothesis is judged to be a good enough explanation. In this manner, the cognitive system may engage in corrective belief revision and noise identification via abductive metareasoning.
We experimentally validate both the abductive reasoning and combined abductive reasoning and metareasoning systems with a software implementation. We explore three intentionally-simplified problem domains: simulated object tracking, aerial tracking, and inference to the best explanation with arbitrary Bayesian networks. These domains are intentionally simplified so that we can clearly identify how performance in these tasks is affected by various parameterizations of the reasoning and metareasoning systems. Our experiments show that (1) abductive reasoning is an effective way of reasoning in these problem domains, and (2) abductive metareasoning brings a significant boost in accuracy and noise identification. These experimental results, plus the system's architectural simplicity, together give strong evidence that abductive metareasoning is an appropriate and effective strategy for a cognitive system to revise its beliefs and arrive at more accurate estimates of its world.