This month I have made the most progress so far. I am definetly on track according to my timeline and actually a little further ahead than I thought I would be at this point. Since I am currently home for thanksgiving break, I enlisted the help of my older brother William. His strong background in math, computer science, and engineering was extremely helpful. As mentioned in my previous post, we used python to create a neural network which gave us a nearly exact approximation of the ENSO values for multiple months. A neural network is a computer system modeled on the human brain and nervous system. A neural network uses a large number of processors arranged in tiers. The first tier receives the raw input data, and adds a 'weight' onto it. Each successive tier receives the output from the previous tier. The last tier produces the output, which in this case, is the ESNO value. I have attached a picture I found on google images to make this clearer. The computer system we used automatically adds a weight and keeps readjusting it until it gets closer and closer to the actual value it is looking for. My last blog post shows a screenshot of that graph. Although this so far can only predict one month in advance, I am trying to figure out a way to predict further ahead. Hopefully, an entire year ahead.
1 Comment
For my assignment, I decided to interview my brother, William, who is in his 3rd year at Dartmouth College. He is an engineering major, and just finished a class in Applied Machine Learning. When I explained my project to him, he immediately identified it as a problem which is well suited to be solved using recurrent neural networks. He used python to create a recurrent neural network that uses the past 24 months to predict the next months ESNO value. He walked me through how to use this software and how recurrent neural networks work (which I will explain more in my next blog post). Although this current version only predicts 1 month in advance, it is possible to adjust it to predict further ahead. However, predicting further ahead will decrease accuracy, which is not ideal. The graph which I have attached below shows predictions for 250 past months. We decided to test past months just to see how close the predicted values were to the actual values. As you can see, our prediction was incredibly close and the graphs are almost identical. I asked William what other things this type of neural network could be used to examine. I was also extremely curious in machine learning, and asked what other types of machine learning algorithms exist. William told me that this type of recurring neural network is commonly used for stock market predictions, and other instances when you want to use past experiences to predict the future. He also shared with me that there are decision trees, support vector machines, and clustering algorithms which are all different types of machine learning that have their own advantages. I hope to be able to learn more about these in the future!
|
Archives
March 2018
Categories |