CCO: A RL-based optimization algorithm

[ilya-kolchinsky]

The currently used B&B optimization algorithm might be extremely slow for large workloads due to its worst case exponential runtime.
A local search-based algorithm (#37) could be a good intermediate solution, however it is based on a meta-heuristic and thus highly sensitive to the solution space topology and neighborhood definition. A more promising approach could be to utilize a reinforcement learning framework.
Implement an optimization algorithm based on one of the well-known reinforcement learning approaches (e.g., DQN, DDPG, TRPO, etc.).