Upload
cecily-anthony
View
223
Download
0
Embed Size (px)
Citation preview
SE367 Project Final Presentation
By:Sujith Thomas
Parimi Krishna Chaitanya
In charge:- Prof Amitabha Mukerjee
To make a neural net learn the rules of Conway’s game of life and predict the next generation of cells.
To identify oscillators and other emergent patterns using recurrent neural networks.
Simple rules of Conway’s game of life
Emergence of complex patterns
Backpropagated neural network
Recurrent neural networks
Training Neural Network to learn the rules of Conway’s game of life
Training a Recurrent Neural Network to detect a repeated pattern.
1. Input vector of size 9
2. Hidden layer has 9 nodes3. Output layer has 1 node
4. We use bias at input and hidden layer
5. Our activation function is sigmoid
6. We update the weights through the backpropagation algorithm
Features of training model
Input vector of size 18 Hidden layer has 18 nodes Output layer has 2 nodes Bias is present at each layer Activation function is Sigmoid We are again updating weights
through backpropagation. In input vector the last 9
dimensions correspond to previous delayed state as shown.
We are using an array to store the previous 12 output states (size may vary later).
The game has cells of 12 rows and 12 columns .
We use a seed of size 3X3 and4X4 to initialize the game.
We use a activation feedback from the output layer with a delay of 12 ticks.
This helps us to detect oscillators with period 1,2,3,4,6.
Till now we have detected still lives and oscillators. Till final demonstration we will show Gliders after
they are recognized. The problem with gliders comes with their property of “Translation”
For solving this we can either use a 4 layer Neural Network or we have a heuristic of re-seeding.
A guide to Recurrent Neural Networks and Backpropagation, Mikael Boden, Halmstad University 2001.
Pattern Classification – Duda, Hart and Stork
Wikipedia – Conway’s Game of Life Implementation of Neural Networks in C -
John Bullinaria, University of Birmingham. http://www.cs.bham.ac.uk/~jxb/NN/nn.html