AlphaGo in a Personal Computer?

A Google Artificial Intelligence (AI) program recently defeated a world-renowned master Go player Lee Sedol three times, in Go matches which has been known as an extremely complex game (Refer to the news article). According to the ‎Nature‬ article written by David Silver et al. (2016), ‎AlphaGo‬ defines not only value networks evaluating board positions but also policy networks choosing moves, and both networks are trained through supervised learning. The AlphaGo uses a distributed system consisting of 1,202 CPUs and 176 GPUs, which couldn't be easily implemented in a small lab. My interest is whether it is feasible to have similar performance with less system resources (for example, even in a personal computer)? I have no idea, but it might be feasible if both networks could be trained continuously using a concept of Collective Intelligence based on Web Cloud (such as Wikipedia). If you are a computer scientist who are interested in how the AlphaGo works, please visit the repository of source codes for AlphaGo.
Post a Comment