The paper [Constructions in combinatorics via neural networks][1] by Adam Zsolt Wagner uses reinforcement learning algorithm and the deep cross-entropy method, to find explicit constructions and counterexamples to several open conjectures in extremal combinatorics and graph theory! 
Let me also mention the paper [Refuting conjectures in extremal combinatorics via linear programming][2] also by Adam Zsolt Wagner.
(I thank Geordie Williamson for telling me about Adam's work.)


  [1]: https://arxiv.org/abs/2104.14516
  [2]: https://arxiv.org/abs/1903.05495