Introducing PlayTwixT.com. Can you beat our Grandmaster AI?

Hi everyone,

We’re excited to share PlayTwixT.com, a new site we’ve been building to help more people discover, play, and enjoy TwixT.

One challenge we’re especially excited to put in front of this community is our Grandmaster AI. So far, no one has beaten it. We’d love to see what happens when more experienced TwixT players take it on. Could you be the first?

If you prefer a human opponent, it’s also a great place to play with friends or test yourself against other players. You can enjoy a casual game or play competitively, with an Elo-based rating system like those used in chess to track your progress and work your way up the rankings.

Our goal is bigger than building a tough computer opponent. We want to support and grow the TwixT community by making the game more accessible, helping new players learn, and giving experienced players more reasons to keep playing. We’re excited about what modern technology and AI can contribute, both as challenging opponents and as tools for understanding this wonderful game.

We appreciate the players, developers, and sites that have kept TwixT going over the years. We hope to add something worthwhile to that community and help introduce a new generation of players to it.

We’d really welcome your feedback. Bring a friend, put the AI to the test, and tell us what works, what needs improvement, and what you’d like to see next.

Give it a try: PlayTwixT.com

Thanks for taking a look. If you hand the Grandmaster its first defeat, please come back and share the game!

Very nice. I got beat by the AI 3 times in a row. Some things missing:

• The AI played L12 and there was no way to swap it. Does it have any concept of the pie rule? This gives the first player a strong advantage.
  
• Does it have any notion of link crossing or link removal? I see no link crossing, and no affordance for link removal.
  
• Crucial diagonals (helpful for eyeballing whether a ladder chase makes it to the edge).
  
• Coordinates along the edges (numbers along the top and bottom, numbers down the sides).
  
• All of the above makes me wonder if the AI has been thoughtfully trained to recognize draws, which is a known flaw in Twixtbot. Twixtbot fails to recognize a possible draw and walks right into it, failing to play obvious moves that would lead to victory instead.
  
• I beat the AI in "Advanced" mode. Apparently, this AI is of the Never Resign philosophy. It should at least resign when columns 2 and 23 are connected, as red pegs are illegal in columns 1 and 24 so there is no way to avoid connection at that point. But it would be better if it would resign at a reasonable point, and not drag the game out unnecessarily. Connected games on Little Golem are an average of 12.27 moves longer than resigned games.
  

Hey Alan, thanks for testing and for the very thoughtful feedback! I removed the pie rule and link removal for simplicity purposes however can definitely add that in there if it's desired. I'll work through your other feedback as well!

I played one game vs. Intermediate. Went first, took center square. I won without difficulty. Many of the moves appeared random - it had no idea how to form a plan at this level. I assume the higher levels are much stronger, especially if it's beating Alan. From this one game, I would estimate about a 1200 rating.

Two more wins at intermediate, and letting it go first did not help. My 1200 might have been too generous. 1100? Not a useful opponent at that level, but I could see it as a useful practice tool for someone learning the game.
Meanwhile, I'm still too scared to try the higher levels. When I do, I'll let you know what happens. Meow!

MisterCat, are you sure "intermediate" is the right level for you? You are currently ranked #157 out of 2047 on LG, according to this app, which works out to 92nd percentile.

I'm not saying the bot is necessarily calibrated well. To me, "Advanced" felt easy, but "Grandmaster" felt nearly impossible, like Twixtbot. I'd like to make the point that last I checked, Twixt has one of the highest rating spreads on LG, despite a relatively small player count. This backs up my intuition that Twixt is a game with a steep initial improvement curve. Casting this in terms of AI bot play, if you want to cover the full range of human Twixt skill levels, I'd say there needs more than 5 skill levels presented, because I'd expect a lot of people to be in between levels, bored at one level but crushed at the next.

Hi Alan. Obviously 'intermediate' is NOT the right level for me if I wished to play any sort of a serious game. That was not what I was doing. I was there to do a favor for the site operator - he wanted people to try it and give feedback. So - I tried it and gave feedback. Sure - I should try the next higher level and see whether I am crushed. Will do - when next I have some free time. mc.

I just defeated advanced with no difficulty, and it went first. It can't see far enough ahead to realize that it is pursuing losing ladder chases. At least at advanced.
No - I have not tried grandmaster yet.
It is NOT my attempt to be hyper critical or belittle the creators of the site and the software. I think it's GREAT (where is Tony the Tiger?) that they did this, and I expect that they will be tweeking the software as more people play and provide feedback! mc.

Thanks for your feedback! The Intermediate and Advanced AI definitely need more work to improve their skill level proportionally. I spent most of my time perfecting Grandmaster so when you give that a shot, please let me know! Thank you!

I pitted the Grandmaster level against Twixtbot. It beat Twixtbot at MCTS=0 (Monte Carlo Tree Search; at 0, it's just neural network first pick), but it lost to Twixtbot at MCTS=100. Twixtbot at MCTS=0 beats me about 9 times out of 10.

First of all, thanks for your hard work!

I agree with Alan about the swap rule, standard rules (link removal and rearrangement) versus PP rules that are in place on this server (I suggest you offer both rulesets as options.) Also guide lines and coordinate labels around the board would make the board easier to read and talk about.

One possible remedy for the huge strength gap between GM and the next step down is row handicapping. Check out Wikipedia for details. This is implemented on the real-time server Game Center. This would require training your bot on each handicap level. Right now, your bot is trained on the smallest handicap, called "move" which means there is no swap allowed. So you should probably retrain the bot to also play the even game with swap, so it can work out for itself what a balanced first move should be. Stronger handicaps would be move+1 row, move+2 rows ... perhaps all the way to 10 rows? This would result in a very granular handicap system. Likely many humans would find a level where they can beat the GM. I have seen players lose even with move+12 rows. I don't recommend you try to let the stronger player move first in this handicap system. You will get plenty of granularity without that.

I got the "beat the GM" badge by choosing Red. I usually lose even with Red.

Thanks David! That's helpful! Also congrats on beating Grandmaster, pretty awesome :)

I think this is you ^

OK, I played Grandmaster. 1st game it went first, took the center square, and beat me easily. 2nd game it STILL went first, STILL took the center square, but I put more effort into my moves and managed a draw! Considering how this thing has been manhandling Alan and David, I'll take the damned draw! Comments: yes, guide lines would be nice as described above; also, it took it's sweet time recognizing the drawn position, but eventually it dawned on it; highlighting the most recently played move would be handy so I don't have to go hunting. Finally, I still don't think it saves a record of games, even though I am registered. If you'd like to see my drawn position, I grabbed a screenshot of it. mc.

Please let us know when all of Alan's recommendations have been implemented. Also, if not too difficult, it would be very desirable to link your games with Alan's commentator site.

I just want to add about swapping: I believe it is better to swap player colors, leaving the first peg right where it is. That is simpler and less confusing than Little Golem's way, which seems to have been designed under the constraint of never changing player colors. If you want an example, you can go to Twixt Commentator https://twixt-commentator.duckdns.org/ and create a blank board, select the swap style 'Swap Player Colors'.

I agree with Alan. I understand that LG is trying to mirror the swapped board, but it would be less confusing to swap the color.

Thanks everyone! I'm taking all this feedback and making updates and will let you all know when it's ready! I appreciate it!

Quite a number of changes have already been implemented. I recommend checking this out to all who have not done so already. mc.

Thanks everyone for giving it a try and sharing such thoughtful feedback! I’ve finished several updates based on your suggestions:

Swap rule: Players now exchange colors, leaving the opening peg exactly where it was.
Rules choice: You can choose Standard rules, including removing and rearranging your own bridges, or keep the fixed-bridges version. Standard also has computer opponents, though their strength is still experimental.
Board readability: Optional coordinates and ladder guides, plus a marker for the latest move.
Draws and resignations: Improved fixed-bridge draw detection, and the AI can offer to resign when a forced finish is proven. This is deliberately cautious, so it won’t recognize every lost position yet.
Saved games and sharing: Signed-in games are saved for replay, and you can create a public replay link to share a game.

Still to come: better balance between the difficulty levels. Row handicapping and integration with Alan’s Commentator are not implemented yet.

Thanks again, Alan, David, MisterCat and Steven. Your feedback has been genuinely helpful, and congratulations to those who have beaten Grandmaster!

Screenshots:
Swap Colors
Bridge Rules
Board View
Sharing Game

Hmm, I'm really not liking the screenshot site. It immediately wants me to download malware. There are better options, like jumpshare or loom, which are based on the freemium model. Anyway...

Overall, good changes, just a few surprises:

The marker for the latest move isn't showing for me while playing against Grandmaster. Its own previous move is still circled before I hit "Confirm move", and then my move still isn't circled, and then Grandmaster makes its next move and circles it. I'm also puzzled that Grandmaster requires confirming every move, but Advanced doesn't.

The ladder guides are wild. Have you looked at the guide lines on Little Golem games? They just go to the inner corners: b2, w2, b23, and w23. That is useful for answering the crucial question of whose wall a ladder chase leads to. Lines in other places are mostly clutter.

I'm surprised "Standard" isn't the default, and not just because "standard" implies "default". Link removal was part of the rules since 1962, and your website bills itself as "a modern tribute to the classic 1962 board game". Standard should be the default, right?

I'm looking forward to your level calibration. Not only because "Advanced" is still too easy for me and "Grandmaster" kicks my butt, but also because it would be nice to have a site I can recommend to friends. As it is, I'm afraid that friends will complain that the first 4 levels are easy, and Grandmaster is just an exercise in helplessly watching the computer make a line across the board. I really think there should be at least one more level between "Advanced" and "Grandmaster" called "Master". Also, without looking, I can't tell you whether "Beginner" comes before "Novice" or vice-versa. They are synonyms to me.

Thanks for the additional feedback! You caught some bugs there that I'm fixing.

1. Everything should have automated End turns so fixing that. We probably need to allow more time in Standard rules for removing and adding bridges.
   
2. The way it's working now, the latest AI move is circled. Then clicking your own peg to replace bridges circles the peg you're interacting with.
   
3. I'll look at that ladder and update!
   
4. Good idea, I'll switch the order of standard and fixed bridges and make standard default.
   
5. Agree, that the first four levels are too easy and I still need to improve lower levels. I'm thinking remove novice and adding Expert in between Advanced and Grandmaster.
   

Thanks again!