A reinforced learning approach to optimal design under model uncertainty