A Band of 100 GMs Couldn’t Beat the Best Neural Network Chess Playing Computer at One Move A Day

In a chess game of the best neural network for chess vs 100 GMs at one move a day, I would pick an A.I. like Alpha Zero to win.

It is the respect for computing power that brings me to side with a neural network such as Alpha Zero. At one move per day Alpha Zero would have something like Fibonacci Series improvement and upgrades until it exhausted all of the possible moves in chess. It could approach infinity in a sense- a Cantorian trans-finite set paradigm- for learning and structuring chess moves making the human grandmasters about as competitive as 100 monkeys typing Shakespeare vs Shakespeare, in writing another Shakespearian play in 1611.

Alpha Zero’s neural network required just four hours to learn to play chess and beat SF 8? Given a 30 move game, and 30 days to compound its own interest, Alpha Zero’s improvement would be greater than that of Magnus Carlsen vs Gioachino Greco over the centuries. Humans don’t learn remotely as fast, nor have the power for simultaneous equations as a state-of-the-art neural network. Formulaic models written by 100 GMs in a day would be comparatively primitive to those invented by the best and brightest neural network.

Computers presently are up to 122.3 petaflops (IBM Summit June 2018). Neural networks including Alpha Zero I would think, may exploiting any number of computers, including super-computers. The power applied to a single chess game or match would be absurd.

One hundred GMs with about any sort of non-electronic approach calculating board positional vectors would have as little of a chance as 100 of Emperor Gao Gui Xiang Gong best suanpan (abacus) wielders versus a supercomputer at calculating numbers of rice grains in theoretical inventories of all possible Universes minus 1.

For GMs to beat an AI neural network, logically they would require comparable processing capability in critical areas. One might anticipate that human reasoning algorithms regarding chess might be modeled by program designers, yet not vice versa.

Simple Alpha Zero

Fritz beat World Champ Vlad Kramnik in 2006 running on a Xeon dual-core 5160 chip at 3 mgz. The chip sold for just $860. It was a very tiny chip in comparison to today’s chips used for desktops such as the 32 core AMD Threadripper chips.

Neural network programming for artificial intelligence has advanced quite a lot since the 1996–97 Deep Blue-Kasparov games. Apparently the A.I. programming was good old brute force rather than sophisticated algorithm. The A.I. designer for Deep Blue discusses that project here… https://www.scientificamerican.com/article/20-years-after-deep-blue-how-ai-has-advanced-since-conquering-chess/

I took a programming and systems analysis course in 1980, though I worked in other fields, so I understand the philosophy behind designing computer algorithms to process data reasonably well. Computers today are orders of magnitude more powerful than those that beat Kasparov and Kramnik. There is a field of study called Big Data that is remarkable to those that know nothing of it. Big Data sweep with neural networks can get all of the information there is on-line about chess rather easily

A.I. neural network designers today can use everything ever put on-line concerning chess and program a network to sort through it and process it in numerous ways to ultimately let it select the best move. Capablanca, Morphy, Pillsbury, Steinitz; every game ever played and noted online would be in its live database. Deep Blue had very limited capacity for that, and it had just two GMs help its design. With the way data is now, the question to me is could the A.I. read and convert into meaningful information every chess book in any language that it could input. It could have the help of thousands of GMs in a sense providing data and it might build chess ‘wisdom’. Those are just programming challenges that probably were already solved. Some have said that IBM may have been using Deep Blue just to test its A.I. Maybe Google and Alpha Zero were doing that too. It was a learning test.

AlphaGo Zero Explained In One Diagram – Applied Data Science – Medium