Actually, convolution is exactly the same pattern search, just trade memory space to process than sequentially. They are exactly the same mathematically, only that CNN filters are trained, instead of easily designed with patterns that can be understood. And in fact I didn’t even use these in the model, it was used for generating questionnaires out of the game records of a player, and find patterns of the highest appearing frequency to ask (or the questionnaires will be way too huge to finish, like thousands of combinations within local patterns), they only need to answer what they’ve played in the past often, and discard those not of their styles. By default, the trained model will be neutral for uncertain patterns. And one of the perceptions I did try was how much influence or territorial a move would be, but in isolation, it is not an easy concept to train, since human perceptions varied a lot and heavily depend on context. The height of the move is easy to detect, but not their combinations. Sometimes a lot of high positions are just a bunch of disconnected stones, not a group, and what a kyu player of different level think of influences are very “questionable” at best. And dan level players are very picky as to context, and they often don’t agree with each other. The question is what constitute a “moyo”, a few stone and a large jump constitute a territory? The early fuseki are easier, but that is just a few moves which can be easily split and some don’t even think they are territorial at all, but just “normal move”.
The fact is we can actually use the territory estimation belonging as a guide to where might be territory wise, but the more complex joseki, and life and death would often mess it up really good. What high level concept we think is concrete is actually pretty fuzzy, if you want to build training corpus for them in practice (you need to make way more training data of all kinds for supervised learning to pick up the variations and separate “noises” from nearby stones, they are not that easily separated inside a CNN). And reinforcement learning or using PPO would need some heavy human-in-the-loop to check the quality of the generated model output to be stable or not. (from the very preliminary test I’ve done, without a very sensitive discriminator model for PPO, the model would only work fine with what it seen before, and not generalize well, and hallucinate a lot for what is their “territory move”, since they don’t actually differentiated “side territory” and “territory in the air”, and often just apply the patterns on the side into the air).
The meta-content slip into the generation by adding keywords into them (black, white, move, liberty, 3-4, 4-4), and trick it to think it is about playing go game, even if the prompt specifically state never.
But sometimes, some go terminologies like “just one liberty remaining” still slip in, which someone who never play Go would even understand the meaning.
Maybe you can also try other open source LLM (like task fine-tuned vicuna), they should be getting stronger and stronger giving some time (if GPT-4 only provides slight improvement, open source models probably aren’t far off).
It should go off-topic a bit less, but it could be interesting to try to make it go off-topic and share the results here (for example, I know you can get it to go off-topic by starting off with a go-related question and then asking about another topic in the same question, e.g. “what is sente and what’s the capital of the Netherlands?”).
