Who /machine learning/ here? l want to train the circles to go

Down 20 right 20 down 2p left 60 down 40
Solved

>>57175100
They have to learn on their OWN. l won't always be there to tell them what way to choose..

Once a circle dies it sends it's position to the rest of the circles, each position that is sended must be avoided by the remaining circles.

>>57175123
That seems like a way of doing it but l want to test neural networks and the only way of trainning the neural network weights l can think of is using a genetic algorithm

>>57175152
Then evaluate how much deep the circles go, you can maximize this value and add a random side movement.

take a look at reinforcement learning, specifically deterministic q-learning, or stochastic q-learning if you want it to be more of a challenge

I have no machine learning experience, but the way I would do it is send every circle on a random path. Whichever goes the farthest, save that path. Follow that path until you're 10 or so units from completing it, then continue the random game

>>57175236
you could also go for ant-colony optimisation for this

Random guessing.
Do you want unlucky circles?

Also, an gif of this would help

File: aCU0PIB.png (6KB, 726x343px) Image search: [Google]

6KB, 726x343px

>>57175205
The thing is that if l maximize how deep the circles go, the deep values from pic related are the same, and it gets bugged there


l don't understand what you mean by this

File: Untitled.png (23KB, 1009x918px) Image search: [Google]

23KB, 1009x918px

easy as fuck

>>57175111

>I don't wanna have a job

>>57175284
(accidentaly pressed enter)

l don't understand what you mean by this

>add a random side movement.

File: 2016-10-21_18-23-38.gif (391KB, 1108x766px) Image search: [Google]

391KB, 1108x766px

>>57175280
>unlucky circles

What's this?

Here's a gif. The velocity of the circles (v.x,v.y) is defined by a NN(basically a matrix multiplication) that takes the position of the circle (x,y) as input and the velocity as output

The matrix from the NN is orginally inicializated with random numbers(weights), so the movement is kind of random at first

>>57175072
make a mating and a mutation function and just keep iterating

>>57175072
just move in all directions a billion times and it will eventually get it

File: Untitled.png (17KB, 744x638px) Image search: [Google]

17KB, 744x638px

>>57175771
This is exactly what l want to do. But how do l decide which circles mate? Giving the circles that are closest to the goal point does not work like pic related shows

>>57175888
Giving the circles that are closest to the goal a higher chance of mating** does n..

>>57175888
>>57175903
alternatively, calculate fitness according to the length of life, and then do some sort of tournament selection and mutation

>>57175320

Not him but the idea of learning how to train it to solve simple problems is a baby step to training it to solve a harder problem.

>>57175888
what if you just made it avoid walls

After some more thinking, I thought, what if you made it every time you hit a wall, record the coordinate value of that spot. Then whenever you were going to hit the wall, it would avoid it. Eventually you would have the whole thing mapped out and it would have to reach the end or endlessly float.

>>57175072
I guess the problem is that circles can't know their objective, they're trying to solve a maze with no info.
You could give them a bunch of checkpoints and calculate the distance from there, or just make them wander randomly and check their life time.

>>57175072
How do I into ML?
I'm doing an AI course right now but have'nt gotten into ML. What intro book would you reccomend?

>>57175442
Circles that are not lucky enough to make it.

Map the room out into discrete steps and implement Q-learning. They will need to take random steps until they reach the goal, then learn how to reach it. This will probably take quite a bit of training.

What are you waiting for, let it go full genetic, you fuckface

You can't hold mother nature, let those circles be free to mate

>>57176301
Do Andrew Ngs Coursera course, it's really good

>>57175072
why don't you simply use a* or something mate

>>57175442
What library did you render it with? Could you share the code?

I'd do it with a genetic algorithm that takes the best n circles in each iteration as parents to the next generation applying strong selective pressure. Or maybe PSO.

>>57175072
You could view this as a dynamics problem, Have a constant downward force (gravity), have walls emit force as a function of their normal vector, and if the circles ever come to rest at the wrong spot, push it in a random direction.

This approach doesn't require machine learning.

>>57175072
HOMEWORK
O
M
E
W
O
R
K