https://www.codingame.com/multiplayer/bot-programming/bit-runner-2048
Send your feedback or ask for help here!
Created by @Illedan,validated by @pb4,@Agade and @eulerscheZahl.
If you have any issues, feel free to ping them.
https://www.codingame.com/multiplayer/bot-programming/bit-runner-2048
Send your feedback or ask for help here!
Created by @Illedan,validated by @pb4,@Agade and @eulerscheZahl.
If you have any issues, feel free to ping them.
Is there a specific way to check if a player is carrying a prisoner? Thank you!
The prisonerId of a Car entity will either be -1 if it has none, or the ID of the prisoner it’s carrying
Okay, thank you!
I’ve been playing around with neuroevolution. I’m messing with the right inputs for the NN. Right now NN knows the car’s position, velocity and angle, ball’s position and velocity and has some sort of sensors of obstacles in front of it. It allowed to be in top Wood 4 so far. Do NN folks have any tips?
So, you managed to get into second in the top league, reCurse is still too good… do you have any tips for a noob struggling to beat the wood 3 boss. Who is still 10 elo better than me… have made steady inroads but suspect I am missing something massive. any tips or a writeup of your progress would be wonderful
.
I switched to search based algorithm. I use smitsimax with NN evaluation. But I think you can do smitsimax with depth 7-9, with just the raw score as evaluation and beat the boss.
Thank for the fast reply. I was using Duct which should be similar, have the depth set to 10, guess my eval sucks… I tried shallower and thought it played worse. I get about 1500 full sims don’t know if this is good or not, will have to try just the score, thought I tried that and didn’t win, but have fixed quiet a few bugs since then.
If I wanted to try a neutral network, what features do you suggest it should have?
The inputs for my network are: score one, score two, who’s leading in case of tie; for each ball its position and velocity; for each car its position, velocity, if it has ball, angle as sin(angle) and cos(angle). Inputs scaled so most features lie in range [-1;1].
Thanks, and you calc this as your eval at each leaf? I don’t see much co-operation between my cars when I don’t assist them in someway
Also I think plain smitsimax might get past boss4, I haven’t gotten past boss3 with that
Yes, +eval for my cars, -eval for enemy cars (I use tanh (-1;1) as output). At first the cars choose quite randomly, but then their moves are more coordinated and cooperated. Aside from UCT I use e-greedy during selection - 5% chance to select random move regardless of score, it seems crucial in smitsimax.
thanks again, still haven’t gotten past the boss, but you tips were worth a 4.5elo boost so thats pretty decent. I’ve gone form basically 0% against the boss to maybe 25%, and smitsimax is working much better!