Using AI to review and learn from your BGA Wingspan games

[Tedoron]

Perhaps I missed it somewhere, but what makes Claude qualified to provide coaching in Wingspan? What makes it an expert? Why are the results to be trusted?

Fair question because you are right Claude alone is not that good. If you would now paste your game log in claude or chatgpt it wouldn't tell so much relevant coaching advice. You get confident-sounding advice that's often wrong about card powers and the competitive meta. Most of the work I did went into fixing this.

I created a database of bird cards with individual strategic fit and tier. I also created a sort of "strategic bible" aggregating a lot of strategy guides and transcripts from the best youtube channels. I tested over 150 of my. games and give a feedback where I saw that the suggestion wasn't proper.
It's knowledge improving is now with every game analysed that receive a feedback from you testers

I am also working with a friend to create a coaching report based on a completetly different technical approach - more similar to the stockfish from chess. More about it soon.

[illy831]

Perhaps I missed it somewhere, but what makes Claude qualified to provide coaching in Wingspan? What makes it an expert? Why are the results to be trusted?

Fair question because you are right Claude alone is not that good. If you would now paste your game log in claude or chatgpt it wouldn't tell so much relevant coaching advice. You get confident-sounding advice that's often wrong about card powers and the competitive meta. Most of the work I did went into fixing this.

I created a database of bird cards with individual strategic fit and tier. I also created a sort of "strategic bible" aggregating a lot of strategy guides and transcripts from the best youtube channels. I tested over 150 of my. games and give a feedback where I saw that the suggestion wasn't proper.
It's knowledge improving is now with every game analysed that receive a feedback from you testers

I am also working with a friend to create a coaching report based on a completetly different technical approach - more similar to the stockfish from chess. More about it soon.

I'm curious why you chose this approach.

It seems to me that Wingspan is a relatively simple game. The possibilities are quite large I suppose but the goal (acquire max point difference) as well as the mechanisms are well and simply defined, and the total number of moves in a game is fixed. You could develop evaluation metrics for moves, including starting hand decisions, that are deeply ground in a human understanding of those mechanics, and validate on a large dataset of games. You could perhaps use RL or Monte Carlo simulations. That way, the output of the product is data-driven and improvements are objective, directed, and informed. (A tech company would probably find some way to market that as "AI-driven", too!)

A personal tier-list/fit-list combined with a token generator to produce natural-language output similar to Youtube strategy guides, iterated on by manual review of small n results...it wouldn't occur to me that this is fit input for the evaluation tasks you describe. How do you overcome GIGO this way? On top of being highly verbose what would it add more than the fit-list/tier-list/strategy guide itself — apart from the distinct possibility of garbling the context and presenting advice that doesn't align at all with the information extracted from the content videos used as input?

[Katzenoma]

Ued it to analyze my latest solo game. https://boardgame.coach/report/0f557c6d ... 5479fc2e89

The negative points feels a little bit generic, but maybe thats because he had to find something in a well played game

I thought to play hermit trush was a good idea against the automa;-) but the AI thinks it was a white power bird and calls it my biggest mistake:-D

can you share the report first clicking on the share button top right? the link you share is only for you, the button share generates a link that then you can share with others

https://boardgame.coach/shared/29e7f6d0 ... 58db73ba17

If you have a tier list already in your db then I would like to see the starting hand discussed in the luck section

[Katzenoma]

In this one the second miss point is:

Biggest miss · Round 2
The Golden Eagle hunt for Northern Harrier failed, gaining nothing that turn; since Golden Eagle's hunts are dice-based and this one whiffed, the alternative of laying eggs (worth ~1-3 VP that turn given your egg capacity) would have been a safer guaranteed return.

must have been a well played game if an unsuccessful hunt (with golden eagle) was one of the biggest fails (+ the explanations don't make sense)

https://boardgame.coach/shared/61fd4969 ... b18511f7c6

[Tedoron]

Perhaps I missed it somewhere, but what makes Claude qualified to provide coaching in Wingspan? What makes it an expert? Why are the results to be trusted?

Fair question because you are right Claude alone is not that good. If you would now paste your game log in claude or chatgpt it wouldn't tell so much relevant coaching advice. You get confident-sounding advice that's often wrong about card powers and the competitive meta. Most of the work I did went into fixing this.

I created a database of bird cards with individual strategic fit and tier. I also created a sort of "strategic bible" aggregating a lot of strategy guides and transcripts from the best youtube channels. I tested over 150 of my. games and give a feedback where I saw that the suggestion wasn't proper.
It's knowledge improving is now with every game analysed that receive a feedback from you testers

I am also working with a friend to create a coaching report based on a completetly different technical approach - more similar to the stockfish from chess. More about it soon.

I'm curious why you chose this approach.

It seems to me that Wingspan is a relatively simple game. The possibilities are quite large I suppose but the goal (acquire max point difference) as well as the mechanisms are well and simply defined, and the total number of moves in a game is fixed. You could develop evaluation metrics for moves, including starting hand decisions, that are deeply ground in a human understanding of those mechanics, and validate on a large dataset of games. You could perhaps use RL or Monte Carlo simulations. That way, the output of the product is data-driven and improvements are objective, directed, and informed. (A tech company would probably find some way to market that as "AI-driven", too!)

A personal tier-list/fit-list combined with a token generator to produce natural-language output similar to Youtube strategy guides, iterated on by manual review of small n results...it wouldn't occur to me that this is fit input for the evaluation tasks you describe. How do you overcome GIGO this way? On top of being highly verbose what would it add more than the fit-list/tier-list/strategy guide itself — apart from the distinct possibility of garbling the context and presenting advice that doesn't align at all with the information extracted from the content videos used as input?

All fair points - I started leveraging and AI LLM API because it is much simpler than the alternatives. I want to validate first if there is any interest in coaching that is made by a machine.

I started working now with a friend on version 2.0 where we use a completely different approach. We are ricreating a simulator of Wingspan where we can train a model with reinforced learning, similar to what the first chess AI did.
It will take a few weeks of works but is a fun project.

[Tedoron]

Ued it to analyze my latest solo game. https://boardgame.coach/report/0f557c6d ... 5479fc2e89

The negative points feels a little bit generic, but maybe thats because he had to find something in a well played game

I thought to play hermit trush was a good idea against the automa;-) but the AI thinks it was a white power bird and calls it my biggest mistake:-D

can you share the report first clicking on the share button top right? the link you share is only for you, the button share generates a link that then you can share with others

https://boardgame.coach/shared/29e7f6d0 ... 58db73ba17

If you have a tier list already in your db then I would like to see the starting hand discussed in the luck section

Yeah - not the best starting hand - let me pass the feedback and improve the prototype.

Cool that you are playing solo, I have never played solo so I didn't cover so well this use case.
It actually always says you are the winner but what is the goal for you to play solo, do you consider a certain high score a win?