Modelling serendipity in a computational context

Serendipity has played a role in many human discoveries, and an improved understanding of serendipity could help bring about breakthroughs in the field of computational discovery. The concept of serendipity has previously been adopted for users' benefit by many subfields of computer science, in work that focused on employing computer systems to support and catalyse a serendipitous experience for the user. In this article, we switch perspectives to focus on artificial systems that catalyse, evaluate and leverage serendipitous occurrences themselves. Since serendipity cannot be generated on demand, our analysis embraces the concept of $\mathit{serendipity}\ \mathit{potential}$. We survey the literature on serendipity and creativity to distil the core common themes, which we then use as a conceptual framework. Specifically, we describe five operational dimensions of systems with serendipity potential: perception of a chance event, attention to salient detail, a focus shift achieved by interest, bridge to a problem, and valuation of the result. The focus shift is a central necessary condition, in which the system reevaluates or recontextualises something that had been given a low evaluation score, and subsequently finds it to be of increased value. We use our framework to analyse several historical systems which included features of serendipity: Mueller's $\mathsf{DAYDREAMER}$, Figueiredo and Campos's $\mathsf{Max}$, Colton's $\mathsf{HR}$, and Pease's $\mathsf{HRL}$. We show that $\mathsf{DAYDREAMER}$ and $\mathsf{HRL}$ manifest all of the features of our framework. We discuss our framework in relationship to existing work on serendipity, surprise, and control in a computing context. We comment on how environmental factors and system features interact when designing for serendipity, and compare the evaluation of serendipity to the evaluation of creativity.