DeepMind has generalized AlohaGo Zero to also work with Chess and Shogi. Learns superhuman performance through self-play within hours. Very nice work.
arxiv.org/abs/1712.01815
BUT... (See next tweet)
arxiv.org/abs/1712.01815
BUT... (See next tweet)
BUT before you write that headline about how these results show that we have general AI through self-learning, remember:
1. Different networks were trained for each game. The Chess network does not play Go.
2. The networks differ not only in training, but in topology.
1. Different networks were trained for each game. The Chess network does not play Go.
2. The networks differ not only in training, but in topology.
This is because the board state of the different games needs to be represented differently. We need different features for Chess compared to Go, because there are more pieces. So each game requires human engineering.
Another thing to remember is that we know the rules of the games and can simulate them very fast. In other words, we have a perfect and lightweight forward model. This is one way in which board games differ from most real-world problems.
Of course, we can learn and/or construct forward models for real-world problems (and video games), and I'm convinced this is crucial for progress in learning policies for interesting problems. But this is far from trivial, much research is needed.
Note that I'm not saying all this in order to downplay the achievement of the DeepMind team. This is excellent work and I'm very impressed. I just want to pre-empt some of the almost inevitable spin that this paper will be subjected to when reported in the news.
In other words, if you happen to be a journalist reading this and wanting to write something about AlphaZero, keep in mind points 2 and 5 of this earlier list of advice:
togelius.blogspot.jp/2017/07/some-a…
togelius.blogspot.jp/2017/07/some-a…
