Hello there,
Is there a way to not have the game save the replays?
I trained my algo for a night and got 5 GB of replays… I dont need them. Yes I could just delete them after each training round, but if there is a way to prevent them to be saved in the first place, that would be neat 
-Jakob
As a personal solution, you can have whatever script is running all those games delete all files in the replay folders after a game. I will talk to the engineer in charge of the CLI, replay generation does seem like something that should be optional.
There is a variable in the game-configs.json that is called replaysave or something, just set it to 0 from 1.
@Jakobimatrix That’s got to be on the order of 1000 games. Impressive! I’m glad someone seems to be making progress with a ML approach. Looking forward to see what it comes up with!
I (let) train now for 4 Days 24/7 … its slow since
a) The game is played in real time so… about 10 to 40 seconds? a game.
b) I train on my old 2 core 5i something Laptop. So I have 4 Matches running at one time.
4 Colonies a 20 algos. Every one against everyone should be 4*(20^2 + 20) / 2 Matches per generation if I am not mistaken. I have now 113 Generations. Sooo ca. 94920 Matches so far.
And I must say, that until the 100st. generation everything was just randoom shi***
But now I see some patterns in the wall building though they dont seem to be effective… yet
There is a variable in the game-configs.json that is called replaysave or something, just set it to 0 from 1.
You are my hero C1Ryan!
Genetic algorithms, woohoo!
I tried this and gave up instantly when I realized how long I would have to let it run 
I hope this eventually converges to something, I’ll be really curious what you turn up with.
If I were to do an ML approach I would have it train on games from decent hand made algos first to kind of seed it into being somewhat sensible then evolve it into whatever. Though that would require using a reinforcement learning approach before switching to your genetic approach probably.
Just my 2cents, I’m also curious what kind of weird strategies could emerge from pure genetic algos.
If I were to do an ML approach I would have it train on games from decent hand made algos first to kind of seed it into being somewhat sensible then evolve it into whatever. Though that would require using a reinforcement learning approach before switching to your genetic approach probably.
Hmm… Id like my algo first learn to not place walls, where already walls are placed or only destroy walls where walls can be destroyed. If that works, your suggestion would be my second step
i am running games localy with the default thing that comes with the starter algo but for some reson it doesnt save replays. does anyone have an idea how i get it to do so?
Replays should be saved to the /replay folder if you run ./scripts/run_match.sh, which is currently working for me locally. If you are using the test_algo scripts, these do not actually play matches or save replays.
If you continue to have issues, give me a step-by-step and ill try to reproduce it