Global AI and Data Science

 View Only

AlphaZero : Promises of AI

By Moloy De posted Fri January 06, 2023 09:58 PM

"AlphaZero and the Knowledge Revolution" By Garry Kasparov, New York, December 2018

The ancient board game of chess has played a significant role in the history of artificial intelligence, if mostly as a chimera. Founding fathers of computation like Alan Turing and Claude Shannon understood that a simple algorithm could play a competent game, famously demonstrated by Turing and David Champernowne’s ‘paper machine’ program in 1948.

Because early computers were so slow, they also imagined that to ever challenge human chess masters a machine would have to approach chess like a human, using selective, knowledge-based algorithms. The brute force examination of every possible move – well into the millions of possibilities after just a few moves – was obviously too slow.

Of course, these luminaries had no way of knowing that computer processing speed would soon begin to increase geometrically with the advent of integrated circuits. Gordon Moore would not postulate his eponymous Law until 1965, twelve years after Turing’s tragic and tragically premature death.

Much as there can be great beauty within the tight constraints of a sonnet or haiku, the limitations of early computers forced programmers to be creative and experimental. They were addressing AI questions that were far larger than the humble game they were attempting to conquer. Can a program learn from its mistakes instead of repeating the same errors every time? Can a form of intuition be instilled into a machine? Is output what matters when determining intelligence, or method? If the human brain is just a very fast computer, what happens when computers become faster than the brain?

As an aside, chess’s role as a symbol in spreading the computational theory of mind (CTM) is an interesting subject. By 1997, the computer-savvy general public wasn’t amazed that Deep Blue could play world champion-level chess. They were amazed that a human could possibly compete with a machine in a pursuit believed by most to be an exercise in calculation. I was credited with super human, ‘computerlike’ abilities when, in fact, human mastery in chess is believed to be based more on pattern recognition and spatial visualization than on calculation or other computer strengths.

Unfortunately, it turned out that the answers to all these profound questions weren’t required to create a machine that would defeat the world chess champion. As early as the late 1970s, the top programs were all built on the same model: Shannon’s ‘Type A’ machine that used brute-force search, the minimax algorithm, and all the speed the day’s CPUs could provide.

The programming community was happy with the rapid progress in strength, but disillusioned by the straightforwardness with which it was obtained. It was as if the goal to build a robot mountain climber to scale Mount Everest was achieved by a giant tank plowing a straight line to the top. Once Ken Thompson’s hardware-based machine Belle reached master level in 1983, the writing was on the wall – even if many of us would stay in denial for another decade.

This isn’t to denigrate the achievement of the Deep Blue team or the generations of brilliant chess programmers and inventors that came before them – only to put it into perspective with the benefit of hindsight. It’s an important lesson that our original visions are often far off the mark when faced with the pragmatic results machines produce. Our intelligent machines don’t have to imitate us to surpass our performance. They don’t have to be perfect to be useful, only better than a human at a particular task, whether it’s playing chess or interpreting cancer scans.

With the human versus machine era closing, in 1998 I created Advanced Chess to investigate the potential of human plus machine. Once again, chess proved to be a handy laboratory for experiments that had much wider applications. The revelation was that a superior coordination process between them was more important than the strength of the human or the machine. This formula obviously has an expiration date in a small, closed system like chess, where machines grind ever onward toward perfection, but an emphasis on improving interfaces and collaborative processes has become conventional wisdom in open, real-world areas like security analysis, investing, and business software.

By 2017, the ratings of the top chess programs were to Magnus Carlsen what Carlsen is to a strong club player. They use giant opening books, terabytes of endgame tablebases, and multi-core CPUs that make your iPhone faster than Deep Blue. It was hard to imagine that my beloved chess had any more to offer in its old role as a cognition laboratory, and AI game programmers had moved on to video games and Go, a more mathematically difficult game for machines to crack.

In fact, it was a Go program that led to chess’s return to the AI spotlight. DeepMind’s AlphaGo moved beyond pure brute force to compete, and beat, the world’s top Go players. Then the bigger surprise, a generic version called AlphaGo Zero, easily surpassed its great predecessor by eschewing embedded human knowledge and teaching itself to play better by playing against itself.

After seeing this important result, I of course had to ask DeepMind’s Demis Hassabis when he was going to turn his machine’s sights on our favorite game. Humans and machines are relatively very bad at Go due to its complexity. Was there much room for improvement in chess? After decades of stuffing as much human chess knowledge as possible into code, would a self-taught algorithm be able to compete with the top traditional programs?

As we now know, and as this book describes in detail, the answer was a resounding ‘Yes.’ AlphaZero dominated the world’s strongest traditional program Stockfish 8 in two matches, despite calculating far fewer positions.

AlphaZero’s strength is impressive, but its method is far more important. AlphaZero isn’t just applying human knowledge and plowing through billions of positions to generate moves – it’s creating its own knowledge first. And, based on its results and my observations, the knowledge it generates for itself is unique and superior. We aren’t just getting faster results the way we do from a calculator. Instead of a postcard from a far-off land, it’s a telescope that has the potential to let us see for ourselves.

And while chess style is hardly of great interest to the AI crowd, I was quite happy to see AlphaZero’s dynamic, sacrificial style. Not only because it mirrored my own, but because it could play like this, and win, against a fearsomely accurate elite program. Instead of grinding chess into dust with tedious and incomprehensible maneuvering, AlphaZero prefers piece activity and attacking chances.

Of course, a self-taught machine has no use for human heuristics like those, although we cannot help but use them ourselves when referring to its play. It doesn’t think in terms of sacrifices or anything else. It is simply playing what works best, just like every other program. The difference is the journey it takes to arrive and the objective superiority of its output. Instead of the prejudices of centuries of human chess knowledge, AlphaZero distills what matters most in a matter of hours.

One of the great Capablanca’s nicknames was ‘the Chess Machine’, reflecting his invincible consistency. Perhaps in the AlphaZero future, someone who plays ‘like a chess machine’ will be thought of as more of an Alekhine, with dazzling sacrifices and a fondness for unbalanced positions!

The impact of such a system in other areas is difficult to overestimate – again with the caveat that chess is a closed system where all information is known, and no new information can be introduced. Still, virtualization of data is a tremendous shortcut. AlphaZero doesn’t need to aboriously analyze millions of human games when it can play more games against itself in a few hours than have ever been recorded in human history.

One implication is for the decline of the preeminence of collecting human-generated data and expertise, especially in closed systems and elements of open systems that can be effectively broken into closed ones. For example, Google’s self-driving cars have recorded millions of miles on the road but billions of miles virtually. Tesla is hiring digital artists and video-game designers to improve their simulated driving environments in which they train their autopilot. And while it’s not a perfect substitute for the real thing, simulations will only get better and their speed will only increase.

As machines become smarter, we will have to overcome our paranoia and prejudices. This is the next phase of human-machine cooperation, to accept that machine knowledge and judgment can be superior to our own. Instead of just using machines as tools – the ‘centaur’ model – the machines become the experts and humans will oversee them – I call it a ‘shepherd’ model.

Machine learning systems like AlphaZero aren’t perfect, and this is perhaps the most difficult prejudice to overcome, our demand that our machines be completely free of error instead of merely superior. For example, studies show that people prefer a human doctor’s diagnosis to a computer’s even if they are told that the machine is more accurate.

This isn’t entirely irrational, as there will be rare cases where AI systems fail to detect exceptions to their rules, while humans can think creatively and adaptively. And we must not underestimate the power of empathy and the importance of human-to-human relationships in every aspect of building a better society. Technology is a means to an end, not an end. Humans are also prone to dogmatic blind spots and a lack of objectivity, so we should take care not to pass our biases on to our digital creations. There is little point in creating an artificial intelligence that only makes our same mistakes faster.

Chess has been shaken to its roots by AlphaZero, but this is only a tiny example of what is to come. Hidebound disciplines like education and medicine will also be shaken, if slowly, by the improved results promised by AI analysis, if we allow them to. Even if AlphaZero once again looks like the end of the road for chess as a machine cognition research subject, I’m very excited to see where it goes from here.

QUESTION I: How one can think of an AI that will beat AlphaZero?
QUESTION II: Does AlphaZero evaluate chess positions?

REFERENCE : AlphaZero Wikipedia, Game Changer: AlphaZero's Groundbreaking Chess Strategies and the Promise of AI