Combinatorial innovation and research strategies: theoretical framework and empirical evidence from two centuries of patent data

I develop a knowledge production function where new ideas are built from combinations of pre- existing elements. Parameters governing the connections between these elements stochastically determine whether a new combination yields a useful idea. Researchers use Bayesian reasoning to update their beliefs about the value of these parameters and thereby improve their selection of viable research projects. The optimal research strategy is a mix of harvesting the ideas that look best, given what researchers currently believe, and performing exploratory research in order to obtain better information about the unknown parameters. Moreover, this model predicts research productivity in any one field declines over time if new elements for combination or new information about underlying parameters are not discovered. I investigate some of these properties using a large dataset, consisting of all US utility patents granted from 1836 to 2012. I use fine-grained technological classifications to show that optimal research in my model is consistent with actual innovation outcomes, and that the model can be used to improve the forecasting of patent activity in different technology classes.