[Computer-go] On proper naming

"Ingo Althöfer" 3-Hirn-Verlag at gmx.de
Thu Mar 8 02:08:28 PST 2018


Hi Dan, hi friends,

> There is actually no randomness in the algorithm, just like AlphaZero's...

but then it does not make sense to call that algorithm "rollout".

******************************************************
In general: when introducing a new name, care should
be taken that the name describes properly what is going on.

* When Bernd Brügmann introduced the name "Monte Carlo Go"
back in 1993, it did very properly describe the behavior
of the algorithm.

* When Rémi Coulom introduced the term "Monte Carlo Tree Search"
back in 2005, this described very well - for more then ten years
of development - what the algorithm was doing. (Although, over 
the years it became obvious that the rollouts had only a small 
degree of randomness in strong engines.)

* When the AlphaGo team used the name "MCTS" to describe AlphaGo's
search without rollouts but with NN-evaluations instead, things
became problematic. AlphaGo still used the tree search part of
MCTS, but no longer in combination with rollouts.

* And when now someone (Dan) looks at alpha-beta with single-node
extensions instead of iterative deepening, "rollouts" is a completely
misleading term.

I can not predict if Dan's approach will be successful. In any
case, he should look for another proper name to describe it.

Ingo.

PS. One example from the history of astronomy: For many centuries
seven planets were known (from Mercury to Saturn). And suddenly Galilei
got a (little) telescope (in 1609/1910) and discoverd four new
planets [sic!] around Jupiter. It took a while until these moons
were not called planets any longer. And, when in 1801, asteroid Ceres
was discovered and in the following years/decades a good handful of other
asteroids, they first were called planets, later planetoids, and
nowadays finally asteroids. So, there the are chances to repair
wrong names - but it costs energy, and in the meantime confusion
has happened.


More information about the Computer-go mailing list