The defeat of 18-time Go World Champion, Lee Sedol, in March 2016, by AlphaGo (AG) marked a landmark achievement in the domain of Artificial Intelligence because the complexity of this ancient Chinese board game and the level of reaction and intuition required, proved it hard for computers to master, up until then.
A year has passed and Deepmind, Google’s AI group has yet another surprise up its sleeve. Christened ‘Alpha Go Zero’ (AGZ), an AI unimaginably powerful, it derived thousands of years of human Go knowledge from scratch, before inventing better moves of its own, all in the space of three days, simply by playing itself.
Despite having the superior advantage of learning from previously played Go games, AG couldn’t muster a single win in the 100 match game against the newly incarnated and improved version of itself. Using ‘reinforcement learning’, the leaner and meaner AI, with nothing but the rules of the game to its name, played itself over and over again, starting from random play and without any supervision or use of human data. Ordaining itself as its own teacher and learning purely from experience, it prided on discovering winning strategies all by itself thereby refining and getting better at its own game.
We’re moving farther away from training a model to imitate the human data we feed it and are instead looking at a model free from human bias and presuppositions. This goal-oriented learning system is able to derive more nuanced optimal conceptions that we’d take an inordinate amount of time for.
AGZ has baffled us by surpassing our mental capacity, arriving at results experts possibly can’t even fathom. The full gamut of possibilities for how an intelligence might behave is simply too vast to be constrained in any meaningful manner. And if we can’t arrive at a principled and scientific account of the how and why, will we ever have control over this smarter-than-human-intelligence supernova?
Source: https://gizmodo.com/stunning-ai-breakthrough-takes-us-one-step-closer-to-th-1819650084
