Circling the Truth : Model Selection Criteria as a Metric of Verisimilitude in Theory Selection

The purpose of this research is to investigate the possibility of using aspects of model selection theory to overcome both a logical problem and an epistemic problem that prevents progress towards the truth to be measured while maintaining a realist approach to science. Karl Popper began such an investigation into the problem of progress in 1963 with an idea of verisimilitude, but his attempts failed to meet his own criteria, the logical and epistemic problems, for a metric of progress. Although philosophers have attempted to fix Popper’s verisimilitude, none have seemed to overcome both criteria yet. My research analyzes the similarities between Predictive Accuracy (PA) and Akaike’s Information Criterion (AIC), parts of model selection theory, and Popper’s criteria for progress. I find that, in ideal data situations, it seems that PA and AIC satisfy both criteria; however, in non-ideal data situations, there are issues that appear. These issues present an interesting dilemma for scientific progress if it turns out our theories are in non-ideal data situations, yet PA and AIC seem to be better overall indicators of scientific progress towards the truth than other attempts at overcoming the problems of Popper’s verisimilitude. One problematic issue when discussing scientific progress is whether or not our current theories have made any progress towards the truth or have just become better predictive tools. There is an intuitive notion that newer theories are truer than older theories because they appear to identify more true causes of a target system. However, it turns out that it is notoriously difficult to provide an analysis of what it means for one theory to be closer to the truth than another theory. The issue is even more pronounced when considering the pessimistic induction: since all of our past theories have been false, it is likely that all of our current will also be false and perhaps our future theories as well. This poses a problem for scientific realism which holds that identifying the true causes of a target system is an important aim of science. While the discovery of new causes that affect target systems do seem to be an important part of scientific progress, it is not clear that increasing the ability to predict the behavior of target systems must always account for more causes known to affect the target system (Forster and Sober 1994). In fact there is some evidence that our best predictive models and theories might not always be our best explanatory models and theories (Goldsby 2013). However, if we want to define progress in realist terms, there needs to be some account of what proximity to the truth is and how newer theories get us closer to the truth. I will refer to these two concerns as the logical problem and the epistemic problem respectively. An early attempt to overcome the logical and epistemic problems was introduced by Karl Popper in his work Conjectures and Refutations. Popper (1963) called his attempt to overcome the two problems verisimilitude. The concept behind verisimilitude is intuitive in nature – a theory is closer to the truth if it makes more true claims and fewer false claims – but his later commentators would point out critical flaws