Has anybody really grasped the significance of AlphaGo Zero?
The news a few days ago about AlphaGo Zero, an algorithm that easily beat the world’s three best players of Go, sometimes called Asian chess and with seemingly simple rules but with almost limitless combinations, didn’t have much impact in the West, where the game is not played much.
The story was originally published by DeepMind, the company acquired by Google in 2014, and then reported in MIT Tech Review and The Atlantic, which focused on explaining the difference between the milestone obtained by AlphaGo Zero and AlphaGo’s January 2016 breakthrough. The articles help understand some of the basic concepts of machine learning and how this discipline will likely evolve.
As I said at the time, AlphaGo’s achievement in beating the best players in the world was to develop deep learning through reinforcement learning: after training the algorithm with every game of Go ever recorded, the result was a machine capable of imitating the best players, able to predict their plays 57% of the time, but usually unable to improve its performance during a game. The next step was to train the machine with new games invented by itself through combinations, resulting in never-before played moves, and therefore unanticipated by humans, and that were able to clinch matches. AlphaGo managed to beat the best human players thanks to previously tested combinations that had never been used in any game, and that could, despite their very low probability, create winning strategies.
So what does AlphaGo Zero bring to the party? It eliminates the need for the first phase, the contribution of Go matches played by humans. To obtain this new algorithm, which has been able to beat the previous AlphaGo by one hundred games to zero, it was split, as its name indicates, from zero. An algorithm with the rules of the game defined, which starts from there to test moves in games against itself. If you have a subscription, you can read it in Nature. After the adequate number of iterations, many millions of games, the algorithm, which has never been fed information about games played by humans, beat its predecessor, an approach that can be applied to many other areas and eliminating the need for large amounts of historical data, as soon as we deal with comparable problems, this is, stable formulations in highly predictable environments and outcomes.
What are we to make of a milestone like this? For managers, it means thinking about what how this can be applied their value chain or day-to-day operations: very broad areas of combinations, but which generate predictable results based on clearly established rules. This is not a universal panacea, but will certainly solve many problems.
It’s time to start asking ourselves what problems we can solve using algorithms that learn from data stored in our transactional files, our CRMs or in our ERPs over the years, and which ones we can solve without that data, simply by defining the rules properly and training the algorithm from a clean combinatorial space, from zero.
This is a potential game changer: in the vast majority of cases, the greatest investment in time and resources in machine learning projects, around 80%, has to do with the collection of data now stored in relational models; with their transformation and preparation. If we can start from zero for some projects, the benefits in terms of cost reduction and increased performance could be considerable, creating real competitive advantages in the process.
Understanding such situations, being able to think in terms of algorithm training and developing awareness about which techniques to use, depending on the situation, are precisely the kind of skills that companies should be developing right now in their managers, instead of being sidelined by apocalyptic predictions, murderous robots and post-work worlds. It’s time to take a leaf out of AlphaGo Zero’s book and get down to work.
(En español, aquí)
