The imitation game: imitation learning for dynamic decisions under uncertainty

Imitation learning (IL) has become a promising approach for solving large-scale combinatorial optimisation problems by learning decision policies from expert demonstrations. However, the notion of a “good” expert remains unclear. This work introduces a taxonomy of experts based on uncertainty modelling, optimality, and interaction with the learner. We propose a generalised DAgger framework supporting multiple and interactive experts. Experiments on a stochastic assignment problem show that stochastic and interactive experts yield superior policies.