Validation report 002: Go Policy Networks

In this report we aim to spot shortcomings of using a convolutional architecture as a Go policy network. By comparing it to an equivalently trained Transformer policy and employing XAI methods such as Ceteris Paribus, we can see where each network under and overperforms. This work points in the direction of further research of Transformer architectures in positional games such as Go, where previously it was believed Convolutions were SoTA.

Our experiments show the limitations of convolutional policy networks. They are great at capturing local features, however limited in attending to global phenomena. In some tasks, like the game of Go, this might be detrimental. Transformers on the other hand are able to seamlessly incorporate both the local and global understanding via its flexible attention mechanism. This points in the direction of future work on implementing Transformers as policy networks for large positional games.

Link to original publication with a model: (see the report)

Recommended citation: Antoni Hanke, Michal Grotkowski. (2024). "Comparative Analysis of Convolutional and Transformer Architectures in Go Policy Networks." Github: ModelOriented/CVE-AI.
Download Paper