In my defense, I originally published this as a throwaway on my Shortform (it's sort of like LessWrong's take on Twitter), but one of the admins wanted me to repost it as a main post. I agree that ideally I would have wanted to playtest this more before drawing attention to it.
That being said, I do suspect that these rules are probably pretty close to achieving the effect I wanted to achieve- the main doubt I have is whether the king's moves work, or if they need to be adjusted to get the balance right, but even there, there's a good chance that it works. I will make sure to update the article when I get a chance to test it out.
Arimaa is a game that was designed to be hard for AI to master. Its rules are quite different from chess, but it has a lot of chess in its DNA. It has since been solved, but it survived 11 years with a $10,000+ prize for the first computer to beat a human champion.
Now that we are in the age of AlphaZero, I think we are close to the bottom of the barrel of ways to design games to hinder AI. I would be surprised if I saw another pure abstract strategy game that favored humans over AI these days, the way Go and Arimaa used to.
This reminds me of David Sirlin's ill-named 'Chess 2' - one notable mechanic from there is that players have stones that they can use to bid when a piece is captured, and if the defender wins the bid, then the capture fails.
I played a good bit of Really Bad Chess when I was first getting into chess. I'm a big fan of Zach Gage's games, I think compared to his other works it's not as platonically good as his other works, but what I love about RBC is that the difficulty of the board is chosen based on your skill level, so when you're a beginner you get positions that are absurdly unbalanced in your favor, but then as you get better you eventually get the tables turned and have to work your way out of a difficult situation. It goes a long way to make chess feel more accessible and enjoyable for someone who doesn't know what they're doing.
That said, Really Bad Chess isn't really a chess variant. It's just orthodox chess with a semi-randomized board layout, but it changes none of the core mechanics, and all the pieces are the same as orthodox chess
I agree that promotions are an important part of what makes Shogi work so well. As I mentioned in the article, I see there are two ways to make drop chess work - either build a game around weak pieces, or make the king more nimble. I think Shogi has perfected the first approach. I love Shogi. I don't want to reinvent Shogi. But there's something that I love about Western Chess: I love knights, bishops, rooks, and queens. I also like the pawn structures it has. I like Western Chess, but I don't love it the way I love Go or Shogi. My goal is to create a version of chess that I love, and has the knights, bishops, rooks, and pawn structures that I love. This is my current best guess at what that looks like.
Author of the linked post here. I've actually done more Go than Chess in the past 6 months - I enjoy it a lot. My dream is to find a game that captures what I love about both Go and Chess. I've seen many examples of how not to make that game, but my mind turns back to that question every now and then.
The best way to solve this is to say that you can't move a piece right back to where it just came from, like the ko rule in Go. As for superko, I think the threefold rule works just fine.
I'm excited to see this. It reminds me of a paper by Cameron Browne (http://ccg.doc.gold.ac.uk/wp-content/uploads/2016/03/browne_...), which does similarly, but using less advanced technology. While the approach was janky in some ways, the game that the system invented, Yavalath, has received some praise for being an interesting and compelling take on the n-in-a-row genre.
I've been dreaming of doing similar with an alpha-type system, and so I look forward to reading this paper.
Edit: After skimming it, it doesn't look like they're going the distance like Browne did. Browne had the computer evolve games that the AI would rate ideally, while the DeepMind paper only has the computer consider a handful of pre-selected variants. I don't anticipate it would have been much extra work to evolve games, and I would have found it much more interesting.
This is already a known variant (modulo the constraint on the second move not relying on the first). Best way to play it is that the white's first turn is only one move, then black starts with the first double move - this is the 12* protocol (first implemented in a game called Connect6), and it practically eliminates any first move advantage