Latest Posts

Topic: Bechmark-play with Record-list

TuStudent
Avatar
Joined: 2015-05-02, 23:30
Posts: 16
Ranking
Pry about Widelands
Posted at: 2019-03-08, 21:49

GunChleoc wrote:

Also, which tree and which bush gets planted is random. Mining also has randomness in it.

Answer from https://wl.widelands.org/forum/topic/4457/?page=1#post-26993

Nordfriese wrote: ( https://wl.widelands.org/forum/topic/4457/?page=1#post-26993 )

So how does random stuff (combat, mainly) get calculated if both computers do it independently?

"Random numbers" are generated by a deterministic algorithm that is so complex that the results appear random to humans. But the entire game logic (minus desync causers) is 100% deterministic.

A bot can calulate those randomness in andvance and will be better than humans. (Just for a specific Bechmark-Play.)


Top Quote
hessenfarmer
Avatar
Joined: 2014-12-11, 23:16
Posts: 458
Ranking
Tribe Member
Location: Bavaria
Posted at: 2019-03-08, 22:17

TuStudent wrote:

Answer from https://wl.widelands.org/forum/topic/4457/?page=1#post-26993

Nordfriese wrote: ( https://wl.widelands.org/forum/topic/4457/?page=1#post-26993 )

So how does random stuff (combat, mainly) get calculated if both computers do it independently?

"Random numbers" are generated by a deterministic algorithm that is so complex that the results appear random to humans. But the entire game logic (minus desync causers) is 100% deterministic.

A bot can calulate those randomness in andvance and will be better than humans. (Just for a specific Bechmark-Play.)

As far as I know neither Nordfriese nor I am wrong. It is deterministic but not reproducible as it is dependent on the initial seed for the random algorithm and this changes as well.
However it would be hard to develop a bot for widelands as widelands isn't linear at all. The right decision always depends on what your opponent is doing. So as the ways an opponent could play a map are very big (in all its variations I think rather infinite) this would lead to a lot of decision paths for a bot. so currently we are trying to teach the AI the rules of the game and getting it to learn to make good decisions in general.
From my perspective this led us to an AI which is doing a lot of stuff better than before while doing some weird things which we should minimize. By this we hope to get a working AI for a variety of tribes and maps.
If you need some challenge though I would recommend to play King Of Nowhere's challenge maps from the download section. They are designed to be pretty hard for single player versus AI matches.


Top Quote
kaputtnik
Avatar
Joined: 2013-02-18, 20:48
Posts: 1626
Ranking
One Elder of Players
Location: Germany
Posted at: 2019-03-09, 15:36

TuStudent wrote:

Is the a map where there is a benchmarkrecord like the most points after 4h for Collector (no opponent)

I think this an interesting idea and can also be used for single player short time games. This idea can also be applied for other things than collectors points:

  • Most planks after 1 hour, probably on a map where tree growing is diffcult
  • Most meals with barbarian tribe after x hours
  • ...

Just a bit brainstorming:

  • Those benchmarks are likely map specific
  • So they are like a scenario, and can be controlled by lua, e.g. exactly stop at 60 minutes game time.
  • There should be a place where to upload the replays/savegames on the website then, ideally automatically analyzed and displayed as a ranked table for each type of benchmark

Top Quote