Perfect replay editor

[dstien]

Some of the ideas I've had with restunts involves automated replay analysis, which, in addition to aid completely automated competition verification, also could help develop perfect AI using genetic algorithms. Unfortunately, none of us have had time to work on restunts in years.

[BonzaiJoe]

the same but with heuristics? hmm.. these should be really good heuristics to complete a run within a month.
some machine learning? i can imagine to train some situations and then apply it. still it would discard some of the best tricks. but really usable with OWOOT rules! the note: that training has to be huge because of many entrance speeds and many track's pieces combinations to learn.

ok so with the whole engine we could have a machine that makes nearly perfect replays with owoot rules. as long as it does not making money it's good only for ruin a game and probably master thesis.

It boils down to designing a smarter Skid, then. In fact, there is actually some prior art related to that specific issue, though it isn't down the computer science alley...

One of the best events in Stunts culture 

[BonzaiJoe]

"but in a nihilist's world there is no "best".  there is only staying home on a cloudy day and lying still on the couch at awkward angles.  in slow motion."

[CTG]

How about just simulating all possible keypress sequences and detecting the fastest finished one?

A track of ~1 minute has about 1200 'moments', and about 64 possible keypress combinations for each moment, so that leaves only about 64^1200 possible replays, . So if you pre-set a max-time that you know the perfect replay should be within, you can just run all the simulations, and the perfect replay should be ready an infinite length of time after the end of the universe.

Once I asked Zak about the possibilty of a program counting the best available replay. He said no - but what if we have the source code? Is there a program being able to produce the perfect replay? Just think it over:

Average length: 80 seconds (1600 units)
3 different buttons can be used in the same time unit (42 combinations)
Possible keyboard combinations in one unit with real meaning: about 20

The program must count 20^1600 possibilities...

When we are already ensured that a perfect replay is impossible we could start about near-perfect replays. That is what I've done.

Btw player could strike at most 3 keys at a moment but I think all 6 keys could be simulated at once.

Why do I feel that Friker owns that program?

[CTG]

The theoretical possibility of creating such a software is a fact. However, the calculation time for a whole replay would be fuckin' long. BUT! What if you take only short, tricky sections to create magic carpets, extreme jumps or simply a perfect 90 degree corner? Usually it's only a 0.5-2 seconds long section and gear switches are quite limited in a certain part of the track (---> less combinations). The buffer parts, such as straight roads do not really require this kind of method - only a few critical sections need that kind of maxout. It would not create a perfect replay, only a perfect trick or corner.

What do we know?
- You know each cars behaviour from .RES files (torque, weight, air drag, etc.).
- You can reproduce your replays in Excel (by knowing the surface parameters) - if you modify some keypress codes in the file, it can regenerate your position on the map. Optimizing requires advanced programming skills.
- You can easily visualize replays, just think on stuntschallenge.

The above mentioned stuff is enough for that - we have some addicts, being able to create something like that. As for me, the software exists. But I don't really care anymore.

[Friker]

...BUT! What if you take only short, tricky sections to create magic carpets, extreme jumps or simply a perfect 90 degree corner? Usually it's only a 0.5-2 seconds long section and gear switches are quite limited in a certain part of the track (---> less combinations). The buffer parts, such as straight roads do not really require this kind of method - only a few critical sections need that kind of maxout. It would not create a perfect replay, only a perfect trick or corner. ...

This is a way I was thinking about. But you will need a measure which can compare which replay is better than others. And that is quite hard with magic carpets.. Perfectly cut corners/loops could be possible.