During the last decade we have witnessed a paradigm shift in the data that may be collected. While performing large-scale randomized experiments has long been prohibitively expensive, the development of gene editing technologies and the possibility of large-scale experiments, for example, using ads on the Internet, has led to a complete change in the type of data that can be obtained at scale. Observational data, even accompanied by strong assumptions, can only paint a partial picture of the underlying causal mechanisms. However, with sufficient interventional data, causal effects may theoretically be learned with much weaker assumptions.
With the floodgates open the challenge is to control the deluge: the main difficulty in these applications lies in the combinatorial explosion of the possible interventions that can be performed. Despite the vast increase in possibilities, it remains, in general, simply not feasible to perform all possible interventions. It is therefore critical to develop a theory of experimental design for causal structure discovery that informs which experiments to run in order to extract as much causal information as possible.
This workshop aims to integrate ideas from optimal experimental design, active learning, policy learning and contextual bandits.
Further details about this workshop will be posted in due course. Enquiries may be sent to the organizers workshop-causality1 [at] lists.simons.berkeley.edu (at this address).