Neural Network Ressources

General
#69

I’d try with 2 hidden layers of 64 nodes each first. That should fit in with just a base64 encoding so you can concentrate on training + forward pass.

If you find it’s too small of a network you can try larger.

To answer your question you can also try base85 or for further gains try (ab)using unicode. Codingame counts characters, not bytes.

1 Like
#70

Assuming c++ code locally and in codingame, unicode characters for codingame · GitHub
This approach should be translatable to python easily I believe.

If I’m counting right, your current net has 124405. With one less hidden layer that would be 84205. That would leave ~16000 characters for other code.

1 Like
#71

the unicode method is genius!
I will try to train a smaller network first and then try the unicode idea later.
Thanks for both of your help!

#72

After a long time I am back in coding game.
I found this:

(D-600/20000): sqrt((nextCPx - x)²+(nextCPy – y)²)-600/20000

angleScalD : ?

angleCrossD : ?

vScalD/1000 : ((nextCPx-x)*vx+(nextCPy-y)*vy)/(SQRT((nextCPx-x)²+(nextCPy-y)²))/1000

vCrossD/1000 : SIN(ACOS(((nextCPx-x)vx+(nextCPy-y)vy)/((SQRT((nextCPx-x)²+(nextCPy-y)²))(SQRT(vx²+vy²))))))(SQRT(vx²+vy²))/1000

nextDirScalD : (((nextCPx-x)(nextnextCPx-nextCPx)+(nextCPy-y)(nextnextCPy-nextCPy))/((SQRT((nextCPx-x)²+(nextCPy-y)²))*(SQRT((nextnextCPx-nextCPx)²+(nextnextCPy-nextCPy)²))))

nextDirCrossD : SIN(ACOS((((nextCPx-x)(nextnextCPx-nextCPx)+(nextCPy-y)(nextnextCPy-nextCPy))/((SQRT((nextCPx-x)²+(nextCPy-y)²))*(SQRT((nextnextCPx-nextCPx)²+(nextnextCPy-nextCPy)²))))))

(distanceCP1CP2-1200)/10000 : (SQRT((nextnextCPx-nextCPx)²+(nextnextCPy-nextCPy)²)-1200)/10000

Someone know how to calculate angleScalD and angleCrossD ?

#73

double angleScalD = dot(vec2ByAngle(angle), vec2(1.0,0.0));
double angleCrossD = cross(vec2ByAngle(angle), vec2(1.0,0.0));

How to have angleNextDir??
double angleScalD = dot(vec2ByAngle(angle), vec2(1.0,0.0));
double angleCrossD = cross(vec2ByAngle(angle), vec2(1.0,0.0));
int vScalD = (int)(normVdot(vec2ByAngle(angleV), vec2(1.0,0.0)));
int vCrossD = (int)(normVcross(vec2ByAngle(angleV), vec2(1.0,0.0)));
double nextDirScalD = dot(vec2ByAngle(angleNextDir), vec2(1.0, 0.0));
double nextDirCrossD = cross(vec2ByAngle(angleNextDir), vec2(1.0, 0.0));

#74

I have now a working NN and know how to translate the input into the NN
But I didnt find how to translate the output of NN into the game , the second term is the thrust but I dont know how to convert the first term into the destination of the pod.
Any ideas ?

#75

In expert mode you can provide angle (-18,18) and thrust. When you provide destination the referee just translates it into angles internally.

#76

I think Pedro1994 is talking about Mad Pod Racing.

#77 
int x = checkpoint[_p1.check_point].x - 3*_p1.speed.x;
int y = checkpoint[_p1.check_point].y - 3*_p1.speed.y;

or 

float anglef = _p1.angletot;
float angleRad = anglef * PI / 180.f;
Vector2 dird = {cos(angleRad)*10000 , sin(angleRad)*10000};
cout << dird.x << " " dird.y << " " << thrust << endl;
#79

You train your NN on all the track ? or on a random checkpoint ? Mine run on all the track , but I don’t know how to converge the NN for all the tracks ?

#81

PLease can you help me ? Sorry to bother you !!! :smiley:

Did you train your NN on all the tracks, or in a random checkpoints?
i use PPO and train on all the tracks , but my NN converge only if train on only one track? How to learn multiple track ?
Thanks !!!

#82

Hi,

I create a batch of matches, where each match use a different race (generated like done by the referee). Of course, in the input of the NN I gave it the checkpoints positions.

Regards,