Artificial Neural Networks
(CS5652): Final Program Competition
Due date: before noon, Jan. 20, 1997
(Please submit to TA at room 228 directly.)
The final program competition's focus is time series prediction.
That is, at time t, we want to use the past values of a time series x
to predict its value at time t+1. More formally, we want to construct
a neural network that takes the input [x(0), x(1), ..., x(t)] to
generate an output N(t+1) that should be as close as possible to
the true value x(t+1).
To proceed, first you need to download the following files:
- blackbox.m: this is the source of the time
series.
- getdata.m: this is the file to collect the
time series x(t).
- mymodel.m: this is the file you need to
modify. Without modification, the file uses x(t-1) to predict
x(t) and it can achieve an RMSE of about 0.03.
Therefore your model
should be able to perform better than 0.03.
- perf.m: this is the file to evaluate your
model's performance.
After you have all the files, run getdata.m to plot the time series
data.
Then run perf.m to get a feeling of how a simple-minded
approach can achieve a not-so-bad RMSE of 0.03.
After this, you can
start constructing your NN model by modifying mymodel.m.
Note that:
- It is impossible to use all the past values of a time sreies
as the input to an NN. Therefore you need to do input
selection by trial and error.
- You need to submit both the report (in HTML format, as described
in our course homepage) and
the program mymodel.m (along with data files needed by
mymodel.m, if any).
- When I test your program, I'll use a different time span for
perf.m.