Re: [Bug-gnubg] Dice Rolls

[Russ Allbery]

JD <address@hidden> writes:

> While you can do a statistical study of the randomness of dice rolls, it
> does not take into account that the engine can see current board, and
> determine that it (the gnubg pseudo-player, which is also me - since I
> selected human against human, and player 0's default name is gnubg). I
> find it very strange that in 4 out of 5 games, where I was playing both
> as player 0 and player 1, with player 0 being gnubg, the rolls for gnubg
> were incredibly winner rolls - totally amazing to me.

> So, I was not playing against the bot, as you say. I was playing human
> against human, and I did not enable the "Dice Manupulation" option at
> all.

> But I can tell you in all honesty that I was doing my best for each roll
> of the dice for both players (0, and 1).

The problem, in a nutshell, is that your sample size is much too small.

Humans (all humans) have a very bad intuition for randomness.  We expect
it to be more uniform than it actually is, when in reality it's quite
clumpy and prone to long patterns of behavior that we think of as not
random.  This is why, for example, many slot machine players will
absolutely swear that there is such a thing as hot and cold machines, or
that a machine is "ready" to pay out, when under the hood the slot
machines have government-certified mathematical random number generators
that don't care in the slightest about whether they've paid off recently
or not.

We're very, very good at finding patterns.  It's what our brains are
adapted to do.  We therefore find patterns even when they don't exist,
possibly since (evolutionarily) it's a survival characteristic to err on
the side of seeing patterns (e.g., predators) where they don't exist
instead of erring on the side of missing patterns that do exist.

In any event, five games is far, far too few to tell you anything of
substance (as you alluded to in your earlier message).  You'd need to do a
more systematic and recorded study across more like 100 games, with some
sort of objective criteria, to be sure that the pattern is statistically
significant.

-- 
Russ Allbery (address@hidden)             <http://www.eyrie.org/~eagle/>