Training a Neural Network using Genetic Algorithm

[@programmingsoftwaredesign7887]

The nodes jumping prior to the hole are the type of people that do a 360 before they kill you.

[@dingalong14]

Wish I got a "good job buddy" when I lasted 2.94 seconds.

[@funduk89]

Your videos will become very popular soon. I am sure. Keep the great work!

[@Djellowman]

Love your videos! These are my favorite kinds of projects. I've made a pong clone that i want to try a genetic algorithm for, using a neural network package i've written in numpy

[@aliyutube]

I wanna learn more.

[@Juni_0220]

Generation 3 will always have a place in my heart

[@numero7mojeangering]

Because your neural network have two inputs and one input, have you tried to plot it to a texture?

[@PezzzasWork]

I haven’t tried but that an very nice idea! I will test it

[@dezm0n679]

Very cool sim, can i ask few questions? 1. How often NN fetch data from game? (Every frame/every 0.01 sec?) 2. How NN fetch data from game? Or its just implemented in game code? Thanks

[@PezzzasWork]

Thanks! 1. The data is fetched every frame (in this case it was at 144Hz) 2. NN inputs are provided in game code since it's fully homemade

[@BeneathTheBrightSky]

Gen three got really good really fast. I'm surprised!

[@ВсеволодВласов-д7ж]

Wow, thats amazing! Can you publish code on github?

[@PezzzasWork]

I will publish it soon

[@ashvinbhuttoo]

@@PezzzasWork cant wait :D

[@Ab-cj6gl]

@@PezzzasWork yeah me too i would love to see your code I'm building a neural network with JS and i just started coding few days ago lol

[@superposition2644]

@@Ab-cj6gl I am also building a neural network, but it's my first one ever and I've already been coding for years and it still isn't working after days of work. It's training as I write this message, and hopefully when it's done it will finally work.

[@Ab-cj6gl]

@@superposition2644 awesome! I'm building sort of a program that generates neural network for me up to this point I'm able to pass in how many number of input + hidden + output layers i just need to figure out how to pass values between layers i figured the first hafl from input layer to the hidden layers but I'm still not sure how to do it between hidden layers in case i had multiple

[@teenspirit1]

This is wonderful thanks for sharing. Have you tried augmenting topologies?

[@revimfadli4666]

Or evolution strategies with dense/cascaded net?

[@SolvingOptimizationProblems]

Very cool simulation. Is that in C or python? Thanks

[@PezzzasWork]

Thanks ! It is written in c++ :)

[@SolvingOptimizationProblems]

@@PezzzasWork ah, OK, thank you very much

[@npr1m991]

Please what is the output of the neural network ? Do they make a prediction for each frame ?

[@Eren_Yeager_is_the_GOAT]

the output is to make the ball jump

[@GodvenG]

Ok but how they capable of "viewing the map" because im pretty sure the is way for indivuals to obtain some sort of information about next wall instead of just relaying on pure timings

[@pradyumnagupta3989]

What is the search space for the network's parameters? ....by which I mean what is the range from which parameters are initialized and mutations are applied.

[@szymoniak75]

Weights, biases would be my guess.

[@psyneur9182]

Assuming fully connected layers and each connection has a weight and bias float, there are 80 parameters to optimize. This is relatively small number - deep neural networks can have millions or billions of parameters. I couldn't find the source for this specific project, but for another one of his NN projects, he initializes all the parameters between -1.0f and 1.0f. The crossover between 2 networks does not introduce new values, and mutate limits values between -10.0f and 10.0f. So the search space for this project is likely [-10, 10] on 80 floats.

[@teenspirit1]

This is genetic algorithm, not grid search.

[@t.swaggit629]

Very cool! Have you tried doing the same thing using gradient descent (the typical approach) instead of genetic algorithms?

[@karanbirsingh1559]

There is no gradient information here

[@revimfadli4666]

@@karanbirsingh1559 there could be, if you setup a reinforcement learning framework

[@ForTheOmnissiah]

What activation function was used in each neuron? Sigmoid?

[@Andy-vp7vl]

How do you do to make all neurals networks without lagging? Just two of them diminish my frame rate frop 60 to 30

[@BlueSky-fj7yi]

Just code a basic one yourself. I get 250 population 60 fps without lag with an old pc.

[@emmettdja]

Do you pass in the positions of the walls?

[@PezzzasWork]

Just the delta x and y to the closest gate

[@emmettdja]

@@PezzzasWork yea

[@johnlemon8268]

How do you combine the weights from two specimens to produce a new one?

[@PezzzasWork]

It's just a basic crossover, nothing amazing here :D

[@johnlemon8268]

@@PezzzasWork So you have two arrays of weights and you just crossover the weights directly? Madlad.

[@PezzzasWork]

@@johnlemon8268 yes exactly, it's quite brutal and I am working on better ways to do so :)

[@depatrify]

@@PezzzasWork Have you considered using CMAES?

[@vast634]

Whats your cost function? Do you evaluate the number of jump-commands as a cost? (expending more energy)

[@PezzzasWork]

Currently the cost function is just the distance traveled. Since jumps provide a quite big impulse they can't spam it too much or they would just crash on the walls

[@vast634]

@@PezzzasWork Any plans for a more complex environment? Like a mario type sidescroller, with actions like jump and forward-movement, and random platforms. Once input could be the distance to the end of the current platform, and two for the height and distance of the next.

[@PezzzasWork]

That could be fun, you gave me an idea! I will try to develop a little framework to do this :)

[@vast634]

@@PezzzasWork You could also make an adversarial scenario, like in a simple shooter game. Where there are two teams, that try to shoot the other team. Agents can move, rotate and shoot. The sensors are distance to a wall/object, distance and color to an agent that is in front. Own team color. Reload-time for the next projectile. Own remaining "health". It would be interesting to see if they can develop some kind of strategy or optimal pattern to find and shoot at enemies.

[@vast634]

@@PezzzasWork Also a good trick, if you get stuck in a local mimimum, is to reduce the mutation rate (and amount) over time. For example if the performance improvement of the best agent plateaus.

[@vcom2327]

Funny, I wrote an algorithm using genetic algorithms to optimze a Neural Network 35 years ago... I believe I was the first to do so.

[@chinglebingle8243]

God? Is that you?

[@sunday-thequant8477]

What framework/ library you use to the graphics? Nice work!

[@PezzzasWork]

I am using SFML www.sfml-dev.org/

[@sunday-thequant8477]

Thanks my bro

[@sunday-thequant8477]

I love your work 100% respect. I'm doing something similar for trading training agents and i'm using python. But i'm training the agents in all the market data and is very slow... Thats why i decides come back to c++ Btw, You know how to use mongo db on c++?

[@eldattackkrossa9886]

@@sunday-thequant8477 there are accelerators for python, check out numba and tensorflow!

[@sunday-thequant8477]

@@eldattackkrossa9886 yes, not enough faster :D

[@rverm1000]

can you do one in python?

[@darkfrei2]

Which activation function are you using for?

[@PezzzasWork]

It's a sigmoid

[@darkfrei2]

@@PezzzasWork the logistic one?

[@FredoCorleone]

Why does everyone solve this with Genetic Algorithms and not Gradient Descent?

[@PezzzasWork]

Because it’s much simpler :)

[@revimfadli4666]

@@PezzzasWork agreed, though I find evolution strategies to be just as simple, despite being kind of a hybrid of GA and SGD

2:24 this is the albert einstein

[@vagishyagnik]

awesome

[@jonathankiendrebeogo6523]

🤩🤩🤩🤩Wahouuuu 🤩🤩🤩🤩🤩🤩

[@s1mo]

When gen3 is better than me lol

[@sedevacantist1]

This is so misleading; “training” “Neural network” “Genetic algorithm” the only thing being trained is the human programmer. There is no neural network, it's just a computer program. A genetic algorithm doesn't exist, it's just an algorithm (program). This is a flawed technology in that it attempts to mimic a brain by an assumed configuration that can produce inherent logic on its own. The human brain has a hardware system and a programmer all in one package. The delusion in this computer programing technique is that it can be designed to have intelligence even in the most rudimentary sense. This is a strong delusion; that with the right program a computer can make a logical decision that it was not programmed to make.

[@25bever]

you're arguing semantics. real life is exactly like that; so calling this 'genetic' really is pretty accurate. You're right that in the end entities are programmed to do certain things, the thing is: human beings are no different.

[@sedevacantist1]

@@25beverI respect your opinion on the use of the word genetics, and give it the most generous interpretation as an opinion, but genetics referrers to genes, what in a commuter is similar to a gene? My position is that intelligence is metaphysical and your claim is that it is mechanical. The materialist believes that an arrangement of logic circuits can produce intelligence, yet they must tell the circuits to open and close with a program. They claim the neural arrangement in the brain is the producer of intelligence, but that just can’t be true. A commuter is a box of switches and none turn on or off on their own, they must be programmed by an intelligence.

[@revimfadli4666]

So by your logic aren't you also misleading since the computer program you said doesn't even exist, it's just a bunch of voltages and electrons which themselves don't exist, it's just a bunch of quantum waves and so on?

[@revimfadli4666]

So what or who is the programmer of human brain? How were neural nets programmed to do stuff? By the programmer manually setting each weight?

[@sedevacantist1]

@@revimfadli4666 I don’t know exactly what you are asking. A computer is a box of switches that are turned off or on according to a program. These neural network videos make it seem as if someone is making a computer with a reconfigured CPU that resembles a human brain, a completely new type of computer. I believe it is just another program running on a regular PC and that it is less efficient and less accurate. I believe the brain and mind are separate things and the mind uses the brain, if it is operating correctly.

[@BearJew840]

I can do better

[@yosuanicolaus]

lol ...

[@EpicVideoGamer7771]

Prove it.

[@grayphantom100]

its useless , change the level and your training is absolete. i dont get ppl are impressed with these crap click baits

[@PezzzasWork]

If you watch closely you’ll see that the map is different each iteration

[@blubaylon]

🤣🤣🤣