digorithm / artificialintelligencealgorithms

Hi there 👋

I'm a founding engineer and team lead at Fuel Labs. I'm very into distributed/decentralized systems, cryptography, programming languages, and computer science in general. You can read more about me on my personal website.

My work spans across a couple of projects:

1. fuels-rs: The Rust SDK for the Fuel's ecosystem. Think of it as the ethers js for the Fuel network;
2. fuels-ts: Similar to fuels-rs, but in Typescript;
3. Sway: A smart contract programming language targeting the FuelVM and its toolchain (e.g. forc);
4. fuel-core: Fuel core client and node;
5. forc-wallet: Fuel's official wallet CLI implementation;
6. LC-3-Rust: An LC-3 Virtual Machine (register-based) implementation in Rust, for learning purposes;
7. Plus, 10+ years of a lot of closed source work spanning across distributed systems, security, and machine learning, see my LinkedIn for more information.

You can reach me on Twitter @ digorithm.

artificialintelligencealgorithms's Issues

Gradient descent (logistic function)

your article on gradient descent is cool ;)
http://rodrigoaraujo.me/articles/2015-01/Making-Your-Machine-Think-Learn-And-Predict-Gradient-Descent-Algorithm-in-Java

i hacked together some sigmoid functions

	public static double sigmoid(double x, double theta0, double theta1) {
		return 1.0D / ( 1.0D + Math.exp(-1.0D * (theta1*x + theta0)));				
	}

	public static double sigmoid(double x) {
		return 1.0D / ( 1.0D + Math.exp(-x));				
	}

and used them as the hypothesis function, the function with multiple parameters is useful for plotting the graph
then i used the data in this wikipedia example
https://en.wikipedia.org/wiki/Logistic_regression#Probability_of_passing_an_exam_versus_hours_of_study

surprisingly it works and converges, i'm able to use a learning rate (alpha) of 0.1!
then every 200 iterations i reduce the learning rate alpha *= 0.9;
i edited codes so that it compute the change in theta (theta - theta_old) and managed to get it to converge to 1e-4
that gives a rather fast convergence in about 1600 iterations
i got the coefficients pretty close to the wikipedia log regression example, it is probably more accurate :D
cheers