Hydroelectric scheduling (HS) problems, so far, have mostly been solved by mathematical programming methods (linear programming, stochastic dual dynamic programming, etc), because it's the only known method that works well to solve sequential decision making with very large action space (~10000 variables and more).
However, these methods need to simplify the model in several ways, including: assuming the cost function is convex, and simplifying the underlying random processes (like assuming they are Markovian).
Both these assumptions are wrong in reality for hydroelectric systems.

We investigate an alternative way to solve these problems, Monte Carlo Tree Search. It became famous for improving significantly the level of computers for the game of Go. We extended its reach to continuous domains.
However, MCTS suffers from some limitations to work on HS problems: it does not do well on very large action spaces, and it struggles with long time horizon.
This is why we worked on two fronts: first, we developped a framework to mix MCTS with existing suboptimal policies (like Linear Programming), in a way that can get the best out of both worlds.
Second, we are working on modifications of the traditional bandit inspired formula that dictates how to spend computing power in the tree of MCTS. Alternative methods to direct the computing effort show promising results on simple problems.