Youtube channel for game analysis

[druzil]

Hi everyone,

I have quite an interest in game analysis using purpose built game engines to provide high-level insight into what constituents good/bad moves. I have created a pipeline for partially automating analysing replays of bga games (Splendor for the moment, but intend to expand to other games over time).

The process it to take a games replay, recreate the game locally and then run the moves through an engine (alpha-zero type for Splendor) and then take the numerical analysis and try to convert this into commentary that is played alongside a video replay of the game (with highlight overlays).

I'm trying to automate most of this but the commentary still needs curating to bring it up to a reasonable level. I then feed the commentary into a voice model for generating the audio - this mostly to save time as recording this manually I found it too time consuming.

Here is an example of the current generation

https://youtu.be/8iC-ojXf20g

I'm trying to continually improve the quality so any feedback is appreciated.

[dschingis27]

run the moves through an engine (alpha-zero type for Splendor)

Your idea sounds very interesting. Maybe you can elaborate about the engine part? Where is the engine coming from, how is it trained or engineered? I am not aware of a general purpose engine for boardgames.

Like, for instance, AlphaZero/AlphaGo were build by letting them play gazillion games of chess or Go against themselves. Have you done the same with Splendor? The engine building seems to me the most difficult part in the process you described in your post. Particularly, how do you test the engine that it is actually good at the game? I mean, asking provocatively, if the engine is not a very good player then why would I listen to its analysis?

[druzil]

The AlphaZero algorithm is generic and can be adapted to many different games. You can find many instances in github of people creating engines for different games using this algorithm. It tends to breakdown for more complex games. Someone tried using it for Terra Mystic and wasn't able to progress very far. There are white-papers that have more advanced algorithms that can play almost all types of games e.g. Player Of Gameshttps://arxiv.org/abs/2112.03178. Unfortunately there doesn't appear to be any implementations available and the paper suggest it would be quite complex to implement.

Splendor isn't as complex as chess from a tactical/strategic point of view and you can train a decent player with self-play on ~<100,000 games.

In terms of strength the only two points of reference I have are:-
1. Playing against it myself
I've played it several times and not had a win. From an elo perspective I'm currently around 400 so, based on the elo formula that puts the engine at a strength of at least ~600 with 95% confidence after 10 straight losses. If I play more and have more losses that elo goes up which I'm fairly confident it would.
2. When analysing games of very high rated players there is a high correlation between the move chosen by the good player and the engine.

All that being said the commentary that is produced is still not necessary of a high quality as it a difficult task to transform values associated to each move to something that then be articulated as "the engine is suggesting that the strategy in this position is x..."