Résumé
In this paper, we tackle innovation diffusion from the perspective of an institution which aims to encourage the adoption of a new product (i.e., an innovation) with mostly social rather than individual benefits. Designing such innovation adoption policies is a very challenging task because of the difficulty to quantify and predict its effect on the behaviors of non-adopters and the exponential size of the space of possible policies. To solve these issues, we propose an approach that uses agent-based modeling to simulate in a credible way the behaviors of possible adopters and (deep) reinforcement learning to efficiently explore the policy search space. An application of our approach is presented for the question of the use of digital technologies in agriculture. Empirical results on this case study validate our scheme and show the potential of our approach to learn effective innovation diffusion policies.