Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Reimagining Chess with AlphaZero

cacm.acm.org

Post reply on HN
Print view View original

91–100 of 116 posts

Re: Reimagining Chess with AlphaZero

#91

Post by Andrew_nenakhov »

kzrdude wrote: post #81
Andrew_nenakhov wrote: post #76

Earlier quoted context omitted.

It doesn't matter, if there is a time increment, there are still rarely any draws. Maybe 1 in 15 or 20 games. Of course, I'm at the very bottom of that 1%, maybe things are somewhat different higher up. And seriously speaking, the more time given, the higher the probability of a draw: games are usually won or lost because of blunders. If you have time to evaluate all moves and play very patiently, and your opponent d…

I see. I found the rapid world championship, and at the top level the number of draws is certainly higher. (And it's played with an increment). See this table for example: https://chess-results.com/tnr600852.aspx?lan=6&art=5&flag=30 highlighting ½ with the search gives an idea of the number of draws.

First placed MC had 7+ 1- 5=, doesn't look like a draw death to me.

Re: Reimagining Chess with AlphaZero

#92

Post by eklitzke »

bgroat wrote: post #90
thom wrote: post #45

I feel compelled to say that, Coronavirus shenanigans aside, Wijk an Zee has just finished and was a fantastic event with many interesting and decisive games. Classical chess does not need fixing, even at the highest level. It’s fun that we can run these experiments though. I’d love to test out some of the variants a friend and I came up with! One of my favourites was starting with all the pieces off the board and a…

When I was in high school I attempted to code a "Fog-of-War" chess where different pieces had different sight ranges. The King had no sight individually, but their range was the sight-range of all pieces within one piece from them (because they could relay information)

chess.com has a fog-of-war chess mode, although it works a bit differently; pieces can only see squares they can attack.

Re: Reimagining Chess with AlphaZero

#93

Post by gareiner »

istinetz wrote: post #46
pests wrote: post #3

Great read. Interesting variations to the rules. Here's an annotated quote: "The nine changes considered in this study are listed in Table 1. No-castling and No-castling (10) involve a full and partial [not allowed during first 10 turns / 20 plies] restriction on the castling rule. Pawn-one-square, Semi-torpedo [can move 2 squares on 2ed and 3rd ranks], Torpedo [can move 2 squares], Pawn-back [up to your 2ed rank, do…

>Do any chess games support these modes or is these possible to play today? Chess.com currently supports most of the modes in the list - except, I think, semi-torpedo, no-castling (10), and pawn-back. Which, to me, are the least exciting mods anyway.

I think Lichess is the best, and I have read that it is only maintained by one person.

Re: Reimagining Chess with AlphaZero

#94

Post by robrenaud »

Is anyone interested in doing this kind of analysis for games that are in development?

I am a SWE with 16 years of experience at Google with lots of practical/applied ML. I am about to be funemployed, and I'd love to incorporate an AI player/analysis into the playtest feedback loop of a board/card game in development.

I've won the Race for the Galaxy tournament 4 times in 7 attempts at the World Boardgaming Championship (average size of field around 80?) and placed in the Dominion tournament a few times. I created rftgstats and councilroom.com (Dominion), which did analysis/modelling on game logs, but didn't have AI players.

Re: Reimagining Chess with AlphaZero

#95

Post by mrob »

terramex wrote: post #64

I'm surprised that so many comments talk about "too many draws in classical chess problem" as just yesterday we just finished Tata Steel that was a terrific tournament with plenty of decisive games (84 out of 182) and even most of the draws were thrilling games. During commentary, GM Naroditsky mentioned that he asked past World Champions Anand and Topalov about influence of Magnus Carlsen on the game and both respon…

>plenty of decisive games (84 out of 182)

So the majority of games were draws. Do you really think that would be accepted without appeal to tradition? If you released a new game with so many draws, people would call it broken. If a minor rule tweak can reduce the draw count without substantially changing the feel of the game, it seems to me the common sense thing to do.

Re: Reimagining Chess with AlphaZero

#96

Post by zamadatix »

ogogmad wrote: post #67

What about "Progressive chess"? White makes 1 move; black makes 2; white makes 3; etc.

Interestingly I think that would be a trivially solvable game with traditional methods (like checkers, connect 4, or tic-tac-toe were) due to the seemingly overwhelming likelyhood the game could always be forced to end in 5ish ply or less. Could be a fun weekend programming challenge to prove it!

Re: Reimagining Chess with AlphaZero

#97

Post by SubiculumCode »

istinetz wrote: post #46
pests wrote: post #3

Great read. Interesting variations to the rules. Here's an annotated quote: "The nine changes considered in this study are listed in Table 1. No-castling and No-castling (10) involve a full and partial [not allowed during first 10 turns / 20 plies] restriction on the castling rule. Pawn-one-square, Semi-torpedo [can move 2 squares on 2ed and 3rd ranks], Torpedo [can move 2 squares], Pawn-back [up to your 2ed rank, do…

>Do any chess games support these modes or is these possible to play today? Chess.com currently supports most of the modes in the list - except, I think, semi-torpedo, no-castling (10), and pawn-back. Which, to me, are the least exciting mods anyway.

No castling? Okay, 1.D4 it is.

Re: Reimagining Chess with AlphaZero

#98

Post by istinetz »

gareiner wrote: post #93
istinetz wrote: post #46

Earlier quoted context omitted.

>Do any chess games support these modes or is these possible to play today? Chess.com currently supports most of the modes in the list - except, I think, semi-torpedo, no-castling (10), and pawn-back. Which, to me, are the least exciting mods anyway.

I think Lichess is the best, and I have read that it is only maintained by one person.

I agree, I use lichess - but the question was, where can we play these modes.

Re: Reimagining Chess with AlphaZero

#99

Post by sorenn111 »

robrenaud wrote: post #94

Is anyone interested in doing this kind of analysis for games that are in development? I am a SWE with 16 years of experience at Google with lots of practical/applied ML. I am about to be funemployed, and I'd love to incorporate an AI player/analysis into the playtest feedback loop of a board/card game in development. I've won the Race for the Galaxy tournament 4 times in 7 attempts at the World Boardgaming Champions…

Start-up/Consulting idea: use ML/RL strategies and tools and help gaming companies improve their AI. In a variety of games, the AI should be a big component of gameplay (all single player games) and ends up being such a liability. I can imagine the litany of issues that prevent this, but I wonder if a company could accumulate some general templates/models/toolkits that could help AI improve for a variety of gaming companies. Most AI in games are more like Stockfish than AlphaZero and the cost to train these models for each of these games is probably prohibitively high. However, I imagine a variety of gaming companies that provide a 1st person game would not mind outsourcing their NPC AI and perhaps a company could find some synergy across games such that the marginal cost of going from game to game could be decreased.

An example game that fits your board game template more or less are turn based games like Civilization. I love Civ games but the AI stupidity is the weakest part of the game. I know RTS games like starcraft are super hard (probably why Deepmind chose to do it) but perhaps turn based games or games with a slower pace and limited action space are doable for a consulting company.

Idk, food for thought, but if you make a billion dollar company I'm saving this post for my records :P

Re: Reimagining Chess with AlphaZero

#100

Post by mlyle »

qzw wrote: post #59
hutrdvnj wrote: post #54

I think what this analysis really shows is that less time lead to less draws. Maybe we could rethink time controls. Classic could maybe be something like 1h+20sec increment for both sides. This would also make it more interesting to watch such an event live.

Don’t rapid and blitz already kind of fill that role? I’m really not sure that having fewer draws is a great metric to guide optimization for classical chess. First, some games that end in draws are actually fantastic, back-and-forth battles that are as exciting as any win. Second, if there are fewer draws, each win becomes less special, especially for the major tournaments. For a cautionary tale, see baseball. They…

I agree that draws and the possibility of draws increases excitement. But I do think top-level chess results in a draw a bit too often: half or more of games in high-level tournaments, and vast majority of world championship games.

Rules that draw 10-20% less often, and remove a little bit of white's advantage, would be nice.

Post reply on HN