Deep learning can't really guarantee to find a global optimum in most nonconvex problems. However, this chess variant is definitely not something that is likely to pose serious issues to modern systems. DeepMind's StarCraft 2 AI does just fine handling vast variation and imperfect information.
> "this chess variant is definitely not something that is likely to pose serious issues to modern systems"
How do you know that? It must be experimented and proven. It's not something that can be handwaved. With machine learning, I would say "not possible until you prove beyond a doubt".
It's a different problem. SC2 is a realtime strategy game, and Kriegspiel is turn-based. The information asymmetry is different in both games, and the goal is different.
In SC2, you can perform realtime monitoring of changes and track them with multiple maneuvers all at once. In kriegspiel you cannot.
Also in SC2, when you see something, you know what it is. You don't know what kind of pieces are where in kriegspiel. Mistaking a queen for a pawn in kriegspiel can be a game-loser.
From a machine learning model’s perspective they are both turn based. In SC2/Dota case one turn is one frame. There is nothing fundamentally different between these and blind chess apart from the possible moves space being vastly smaller
As noted above, the other differences are that you don't know what type of piece you've located (if you've even managed to locate one).
So taking your analogy, and saying an AlphaStar game lasts 5 minutes, at 60 fps, thats 18000 frames. Kriegspiel games average about 50 moves total (or 100 half moves).
So lets tally these up: SC2 has 18000 frames where on visible turf, full knowledge of occupant enemy pieces are available. Kriegspiel has 100 frames where there is no such concept as visible turf of enemy pieces.
Deduction of if and which piece is occupying a square is the only way to have knowledge of the enemy, and the probability for the deduction to be correct drops significantly after each move.
With such a small amount of information available, the problem becomes significantly different.
The key to successful machine learning is available data. There is so little data that the model you are training will collapse or overfit quite quickly and become useless.
The first versions used what you might call 'camerahack', it just sees the entire map at once but as the player himself would see (i.e. fog of war included). The latest version has camera control, only seeing one screen at a time like humans.
