Symbolic Function Learner with Neural Networks

a new deep learning approach to SFL using NN with a discretized softmax-function and an Operator-function P(x[2,4,71..], operators[+,*,/, sin, cos...]), which applies to each node of expression tree and creates the the inputs to the NN.

Overview

This repository contains a neural network implementation for learning symbolic mathematical functions from data. The main architecture is based on a binary tree recurrent neural network (BinaryTreeRNN) which can represent and learn various mathematical expressions.

Features

• BinaryTreeRNN: The main neural network architecture to represent and learn mathematical expressions in a tree-like structure.
• Training and Inference: The network can be trained with given datasets to infer the mathematical expressions.
• Visualization: Ability to visualize the learned expressions compared to the ground truth.

Usage

1. Prepare your data. An example dataset is provided with x1, x2, and x3. You can use them or replace them with your own datasets.
2. Initialize your network:

   net = BinaryTreeRNN()

3. Train the model with your data:

   dataset_x = torch.stack((x1,), dim=1)  # Example for using single feature x1.
   dataset_y = YOUR_TARGET_DATA  # Replace with your target data.
   net.learn(dataset_x, dataset_y, num_layers, training_steps_by_standard_softmax, training_steps_by_softmax_prime, lr, lambda_L1)

4. Visualize the results using the expression_plotter function if needed.

⚠ CAUTION: If you want dataset_x to contain more than one feature (like x1 and x2), comment out the plotting section on line 116.

Contributions

Feel free to submit pull requests, open issues, and share feedback to enhance and expand the functionality of this symbolic function learner.