Beyond AlphaZero

For the last two months, I've used this blog's Friday post to wander around topics relevant to AlphaZero:-

• 2019-07-19: Interpreting AlphaZero • 'What can we learn from black box models? '
• 2019-07-26: AlphaZero's Zeros • Engine evaluations of '0.00'
• 2019-08-02: Winning Percentage to Centipawns
• 2019-08-09: 'Game Changer' PGN
• 2019-08-16: The Value of Deep Learning
• 2019-08-23: A Chess Epiphany • 'Would Demis Hassabis have been more useful to society as a professional chess player?'
• 2019-08-30: Playing Without a Horizon
• 2019-09-06: 'Game Changer' PGN Explored

As everyone knows by now, although AlphaZero never played in a neutral, supervised competition, it was quickly copied by Leela Chess Zero (aka Lc0, LCZero), which improved rapidly to become one of the strongest chess engines in the world. Leela has since been joined by other AI/NN engines, which are also making a mark in the engine-vs-engine competitions that I've been following since the beginning of the year. Not too long ago, in TCEC S16 Qualification; CCC9 Bonuses (July 2019), I reported,

The first stage of S16, the Qualification League, is well underway. Two neural network (NN) engine newcomers, ScorpioNN and Stoofvlees, are currently leading, while a third NN, ChessFighterNN, is fighting to stay out of last place. A Leela clone running on CPUs instead of GPUS [LCZeroCPU], is in third place. TCEC chat commands point to resources describing the three NNs [see links].

ScorpioNN and Stoofvlees first survived the qualification stage, then the two league stages (finishing ahead of another NN engine, FireNNb, along the way), and are currently playing in the TCEC Premier Division. There they were joined by two other NN engines, LCZero and AllieStein, along with four non-NN engines: Stockfish, Houdini, Komodo, and KomodoMCTS.

As for the other engine-vs-engine event, the CCC, four NN engines competed in the qualification stage. I reported last week in TCEC S16 L1 Finished; CCC10 S1 in Trouble, that the tournament did not go smoothly. In spite of the difficulties, all four NN engines -- Lc0, Stoofvlees, DarkQueen, and Leelenstein -- eventually qualified to the next stage.

The upshot of this activity is that the NN engines and the non-NN engines are roughly equal in strength. Will this continue or will one of the two technologies soon prove dominant?