But only actual experimentation will tell us if AlphaZero has left a meaningful amount of the chess space unexplored. Chess, like all games, boils down to two fundamental components:
- Explore the space of possible moves
- Evaluate the "goodness" of each such move
AlphaZero uses a technique called Monte Carlo Tree Search to perform 1, and Deep Convolutional Neural Network to perform 2. And Bob's your uncle! Ok, that is a gross oversimplification, but let me explain briefly how each piece works.
Games have a set of states and a set of moves which transition between the states. Since they usually have a single start state, and for simplicity, we usually model this set of states as a tree (rather than a graph). "Looking ahead" just means traversing layers down the "game tree". For a game like chess, which a high branching factor, you end up with a tremendous number of states in only a few plies, so much work goes into avoiding as many state nodes as possible (by trying to identify obvious losers and bailing on that path or following strong paths preferentially). MCTS goes for depth over breadth by "playing out" a particular move very deeply into the game. Since there are a limited number of moves for which one can perform this computationally expensive operation, it does so for only a random subset of possible moves. However, the subset is not chosen with a uniform distribution. The moves which are explored can be weighted by any heuristic you like. More about that later.
If you're looking at a winning state, it is easy to give an evaluation. If you're not looking at a winning state, you have to decide if it's "good" or "bad". This is where the neural network comes in. Right after you make a move, it's hard to tell whether the move had a positive or negative effect on your winning potential. But it's much easier to determine after the game is over. So when the game is complete, you now have a win/loss signal for all the board states in that game. This is now trivial training input for deciding whether each of those board states was "good" or "bad" for you, and you can backpropagate that signal through the neural network which gives you the "good/bad" signal for each board state. This is the "deep learning" portion of AlphaZero.
Because this is the stateful portion of AlphaZero, you could also say this is where the "knowledge" is, albeit in a rather abstract form which isn't really accessible to inspection or query.
If we assume that AlphaZero starts out by making uniformly random choices during MCTS, then it is fair to say that it explores the chess space without any particular bias. The DCNN may inform the MCTS layer, causing it to follow "good" paths rather than "bad" ones, but from what I've read, it prefers to follow "unknown" paths rather than "known". Basically, AlphaZero uses every game to increase its knowledge about chess rather than just trying to win the current game. Because it focuses on covering the space rather than just greedily following the most promising path, it learns much more broadly than chess engines with a non-stateful search algorithm, which must use their search cycles as economically as possible.
For this reason, one should suspect that AlphaZero is not leaving a lot of promising paths out of its repertoire. And, as others have noted, it's style is already quite different from most grandmasters, and it continues to teach us new things. Even so, there are free parameters in the AlphaZero design, which you may tune to extract different outcomes.
If you were to create a "most alien AlphaZero", you could take the basic architecture, and then train it on human games. However, this training would build a secondary network called the Human Predictor. The goal of the HP network would simply be to predict the moves most likely to be played by a human for a particular board state.
Now, when you execute the MCTS phase, instead of preferring "moves I haven't explored before", you prefer "moves which humans wouldn't make", leaving all the rest the same as AlphaZero. This should basically implement what you are looking for. My guess is that the play will end up looking quite similar to AlphaZero in areas where AZ already plays in an alien style, and will be inferior to AZ in games where AZ overlaps more with "human style".
The reason should be obvious: AZ has no restrictions on the kinds of moves it makes. When it plays differently from humans, it isn't because it's trying to. It's because it found, through exhaustive trial and error, that those moves are simply better. And not just better against humans, either: better even against itself! It's possible that one could train AZ to beat humans even more efficiently by using the HP module described above to model the counter-moves that the opponent is likely to make during the MCTS phase. Such a modified AZ may find that "pure" AZ is overly conservative because it gives its human opponents too much credit.
It's also possible to train a "Stockfish killer" variant by simply using Stockfish itself to provide the counter-moves during the MCTS phase. You would need to do this strictly during the learning phase, because you wouldn't have the computes to do this during a live tourney, but effectively, the DCNN would learn to play exclusively against Stockfish, rather than AZ, and skip any counter-play that a different engine might provide. Then, during a tourney, the MCTS could revert to the standard algorithm, since the DCNN has already captured the essence of Stockfish and stuffed it in a bottle.
Anyway, sounds like a fun experiment. Good luck!