CodinGame challenges are basically extremely large decision trees, where your program is trying to find the node with the highest possible score. Greedy algorithms don't work because picking a locally optimal path (such as the highest-value child of the current node) can lead to a globally suboptimal result. Nor do any other straightforward algorithmic approaches, such as dynamic programming.
I'm wondering what's the relevant CS theory for this kind of problems. I'm not even sure what to call them, except maybe "non-linear optimization problems".
I'm also interested in approaches to solve them. Would some sort of machine-learning work? It seems like it should, since the ML approach of AlphaStar was successful in tackling the Starcraft game, which seems to be the same class of a problem.
An example of a CodinGame challenge is their Fall Challenge 2020 which essentially is writing an agent to play the board game Spice Road against another agent.