Optimization and control applications through repeated interventions based on structural causal models

Abstract

In many applications of causal inference, we finally aim to optimize a specific metrics by intervention on a manipulatable variable. This optimization is done through the estimation of the intervention effect on the metrics for virtual intervention, but many conventional methods assume that such an intervention operation is performed only once. However, actually if this intervention operation is applied to a real system, errors often remain for the optimal value of the metrics. To overcome this problem, we propose a causal optimization framework, based on multiple interventions, in which errors are absorbed by repeating interventions and the metrics is converged to the desired value. This is also expected to be applied to the control of dynamical systems. Experiments using systematically generated data were conducted to evaluate the properties of proposed method.