Re: [Toss-devel] Payoff discounting

[Lukasz K. <luk...@gm...>]

> Wow. It works.

I can confirm (even though it needs 7, not 6 iterations for me
in the tic tac toe example, but that's not a problem for now).

> I mean, this form of discounting is obviously better than exponential
> discounting. Hutter in his "Universal AI" has a discussion on
> horizons. The reasons for our horizon are practical and human-like:
> intelligently greedy (high resolution for the short term) and
> intelligently persistent (low resolution for time distinctions in the
> long term).

I'm not really sure about the role of horizons for us: in reachability
games you normally want them and this was very obvious for our
entanglement tests (this is true reachability). Have a look at a hint
run of entanglement now - it is uncomparable to what was before!
But besides their existence, we might have to just experiment with
what is the best one. Anyway - I like what we have now a lot!
Will start a hint run of gomoku or breakthrough to get a feeling
how these are played now.

Lukasz