Browsing by Author "Nurmi, Miska Juhani"

Objectives The purpose of this thesis is to consider what the cognitive models of online causal learning are and what they have to offer for the interactive AI approach. In this thesis, an interactive AI system is considered one that focuses on understanding and collaborating with a human user and which can therefore benefit from cognitive models. The general overview of the models is given by replicating some of the computational results of Bramley et al. (2017) which explored cognitive models for online causal learning. The earlier paper contained four models on how people might learn their causal beliefs, and five models on how people might choose where they place their tests, also known as interventions. Thesis also discusses the implications that the replicated models have for interactive AI, both by considering how these models could be better extended into the interactive AI framework, but also by considering a simple AI based system that could make use of such models. Replication The replication was done by reimplementing the original models of Bramley et al. in R and by reproducing the corresponding figures. Out of the four models used for causal belief updating, two were successfully replicated so that the results corresponded to the original paper. It is not certain why the two other models could not be replicated, and the task is left open for future work. Out of the five intervention choice models, four were implemented and three successfully replicated. One of the models was very close to the original results, but this thesis could not conclude whether it fully reproduces the original results. Implications The simple AI model proposed in this performed poorly but was able to show that in theory, an interactive AI system that incorporates such a model might be feasible in the future with further development. Some recommendations to better extend the replicated models into the interactive AI framework were made. Main recommendations were that a better model on how people might choose where they focus their local attention is needed. Furthermore, it should be ensured that the models approximate human behaviour in larger graphs as well.