Re: Bag of Tricks: Choosing actions via the matching law

Falk Hueffner <[EMAIL PROTECTED]> wrote:
> Probably that means I'm not intelligent, but I would cast stonestorm
> 100% of the time.

Doesn't mean your not intelligent, just not human ;-)

The above stuff (law? theory? pseudocode?) is from the matching law
which is an empirical phenomena that psychologists have noticed. No
one knows why people (and other aninals) act this way, they just seem
to

There's also been a bunch of studies where people say what their
strategies are and then they are monitored and it turns out what most
people think they'd do is wrong. My theory - unlike in computers,
where memory is readable and writable, human brains are write-only
memory. Data is stored in a format where you can't just get the data
back out again. So we have all these test and introspection and
guessing methods where we try to articulate what we feel, believe,
would do, etc. and it doesn't always match

Now, it's entirely possible i screwed up when trying to apply this to
games. Maybe it works differently in this context. The one example of
humans the text gave was basketball players and how they decide
whether to shoot 2-pt or 3-pt shots. People don't always shoot 2 point
shots or always shoot 3 point shots. They mix it up. You'd think they
mix it up for all sorts of complex reasons like range to defenders,
that sort of thing, but empircally over a large enough sample size
(whatever that is), it follows the matching law. Freaked a lot of
people out when they published that


On a completely different note, here's a funny story my animal
learning teacher told us. He told us that rats are smarter than
undergrads. Put them both in an environment with two buttons. One
button randomly gives a reward (food pellet or nickel) 30% of the
time, the other 5% of the time. They then tested each to see what
would happen

The rat, having a very small brain, experimented but once it got the
hang of it decided to only push button A. The undergrads, however,
decided in their heads that the reward was not random, it was based on
a pattern and so they kept switching from key to key in various
patterns they "learned". In the end, the rat got rewarded 30% of the
time and the undergrads got closer to 10%

i don't know much about it, but my guess is that it has to do with
n-gram length and brain size. An n-gram is a series of n items that go
together, In IR theory, "united states of america" is a 4-word term
that should be treated as one word (gestalt). Outside of text
processing and search engines, you still have the concept of a
sequence of actions or things that can signal something. Maybe a rat's
little brain only stores sequences of one move. In that case, pressing
button A gives the best results. Maybe humans can look for more
complex patterns, say 5-7 actions in length, so they stare at the two
buttons and come up with patterns like "ABBBAB". The downside is, they
overcomplicate it and miss the shorter patterns that work best. In
essence, our intelligence and ability to handle complexity (for AI
folk, we create a larger "state space") actually screws us up. Note
that "AAAAA" might have gotten better results but because of time
limitations and the size of the state space, they might never have
investigated that

Don't know if that helps, but at least i think it's interesting :)

-baylor