Lab 2 MSc: Time Series Prediction with GP
NB! This coursework is only compulsory for MSc students taking the 20cr module. We released a different Lab 2 with an earlier deadline for UG students taking the 20cr module.
You need to implement one program that solves Exercises 1-3 using any programming language. In Exercise 5, you will run a set of experiments and describe the result using plots and a short
discussion. (In the following, replace abc123 with your username.) You need to submit one zip file with the name niso3-abc123.zip . The zip file should contain one directory named niso3-abc123
containing the following files:
the source code for your program
a Dockerfile (see the appendix for instructions)
a PDF file for Exercises 4 and 5
In this lab, we will do a simple form of time series prediction. We assume that we are given some historical data, (e.g. bitcoin prices for each day over a year), and need to predict the next value in
the time series (e.g., tomorrow’s bitcoin value).
We formulate the problem as a regression problem. The training data consists of a set of m input vectors X = ( x (0) ; : : : ; x ( m - 1) ) representing historical data, and a set of m output values
Y = ( x (0) ; : : : ; x ( m - 1) ), where for each 0 ≤ j ≤ m - 1, x ( j ) 2 R n and y ( j ) 2 R . We will use genetic
programming to evolve a prediction model f : R n ! R , such that f ( x ( j ) ) ≈ y ( j ) .
Candidate solutions, i.e. programs, will be represented as expressions , where each expression evaluates to a value, which is considered the output of the program. When evaluating an expression,
we assume that we are given a current input vector x = ( x 0 ; : : : ; x n - 1 ) 2 R n . Expressions and evaluations are defined recursively. Any floating number is an expression which evaluates to the value
of the number. If e 1 ; e 2 ; e 3 , and e 4 are expressions which evaluate to v 1 ; v 2 ; v 3 and v 4 respectively, then the following are also expressions
(add e 1 e 2 ) is addition which evaluates to v 1 + v 2 , e.g. (add 1 2) ≡ 3
(sub e 1 e 2 ) is subtraction which evaluates to v 1 - v 2 , e.g. (sub 2 1) ≡ 1
(mul e 1 e 2 ) is multiplication which evaluates to v 1 v 2 , e.g. (mul 2 1) ≡ 2
(div e 1 e 2 ) is division which evaluates to v 1 =v 2 if v 2 6 = 0 and 0 otherwise, e.g., (div 4 2) ≡ 2 ;
and (div 4 0) ≡ 0 ;
(pow e 1 e 2 ) is power which evaluates to v 1 v 2 , e.g., (pow 2 3) ≡ 8
(sqrt e 1 ) is the square root which evaluates to p v 1 , e.g. (sqrt 4) ≡ 2
(log e 1 ) is the logarithm base 2 which evaluates to log( v 1 ), e.g. (log 8) ≡ 3
(exp e 1 ) is the exponential function which evaluates to e v 1 , e.g. (exp 2) ≡ e 2 ≈ 7 : 39
(max e 1 e 2 ) is the maximum which evaluates to max( v 1 ; v 2 ), e.g., (max 1 2) ≡ 2
(ifleq e 1 e 2 e 3 e 4 ) is a branching statement which evaluates to v 3 if v 1 ≤ v 2 , otherwise the expression evaluates to v 4 e.g. (ifleq 1 2 3 4) ≡ 3 and (ifleq 2 1 3 4) ≡ 4
(data e 1 ) is the j -th element x j of the input, where j ≡ jb v 1 cj mod n .
(diff e 1 e 2 ) is the difference x k - x ‘ where k ≡ jb v 1 cj mod n and ‘ ≡ jb v 2 cj mod n

(avg e 1 e 2 ) is the average

j k - 1 ‘ j P

max( t =min( k;‘k;‘ ) - )1 x t where k ≡ jb v 1 cj mod n and ‘ ≡ jb v 2 cj
mod n
In all cases where the mathematical value of an expression is undefined or not a real number (e.g., p- 1, 1 = 0 or (avg 1 1) ), the expression should evaluate to 0.
We can build large expressions from the recursive definitions. For example, the expression
(add (mul 2 3) (log 4))
evaluates to
2 · 3 + log(4) = 6 + 2 = 8 :
To evaluate the fitness of an expression e on a training data ( X ; Y ) of size m , we use the mean
square error
f ( e ) = 1
m
m - 1
X j
=0
y ( j ) - e ( x ( j ) ) 2 ;
where e ( x ( j ) ) is the value of the expression e when evaluated on the input vector x ( j ) .
Exercise 1. (30 % of the marks)
Implement a routine to parse and evaluate expressions. You can assume that the input describes a syntactically correct expression. Hint: Make use of a library for parsing s-expressions 1 , and ensure
that you evaluate expressions exactly as specified on page 2.
Input arguments:
-expr an expression
-n the dimension of the input vector n
-x the input vector
Output:
the value of the expression
Example:
[pkl@phi ocamlec]$ niso_lab3 -question 1 -n 1 -x "1.0"
-expr "(mul (add 1 2) (log 8))"
9.0
[pkl@phi ocamlec]$ niso_lab3 -question 1 -n 2 -x "1.0 2.0"
-expr "(max (data 0) (data 1))"
2.0
Exercise 2. (10 % of the marks) Implement a routine which computes the fitness of an expression given a training data set.
Input arguments:
-expr an expression
-n the dimension of the input vector
-m the size of the training data ( X ; Y )
-data the name of a file containing the training data in the form of m lines, where each line
contains n + 1 values separated by tab characters. The first n elements in a line represents an input vector x , and the last element in a line represents the output value y .
Output:
The fitness of the expression, given the data.
1 See e.g. implementations here http://rosettacode.org/wiki/S-Expressions
Exercise 3. (30 % of the marks)
Design a genetic programming algorithm to do time series forecasting. You can use any genetic
operators and selection mechanism you find suitable.
Input arguments:
-lambda population size
-n the dimension of the input vector
-m the size of the training data ( X ; Y )
-data the name of a file containing training data in the form of m lines, where each line
contains n + 1 values separated by tab characters. The first n elements in a line represents an input vector x , and the last element in a line represents the output value y .
-time budget the number of seconds to run the algorithm
Output:
The fittest expression found within the time budget.
Exercise 4. (10 % of the marks)
Describe your algorithm from Exercise 3 in the form of pseudo-code. The pseudo-code should be sufficiently detailed to allow an exact re-implementation.
Exercise 5. (20 % of the marks)
In this final task, you should try to determine parameter settings for your algorithm which lead to as fit expressions as possible.
Your algorithm is likely to have several parameters, such as the population size, mutation rates, selection mechanism, and other mechanisms components, such as diversity mechanisms.
Choose parameters which you think are essential for the behaviour of your algorithm. Run a set of experiments to determine the impact of these parameters on the solution quality. For each parameter
setting, run 100 repetitions, and plot box plots of the fittest solution found within the time budget.
A. Docker Howto
Follow these steps exactly to build, test, save, and submit your Docker image. Please replace abc123 in the text below with your username.
1. Install Docker CE on your machine from the following website:
https://www.docker.com/community-edition
2. Copy the PDF file from Exercises 4 and 5 all required source files, and/or bytecode to an
empty directory named niso2-abc123 (where you replace abc123 with your username).
mkdir niso2 - abc123
cd niso2 - abc123 /
cp ../ exercise . pdf .
cp ../ abc123 .py .
3. Create a text file Dockerfile file in the same directory, following the instructions below.
# Do not change the following line . It specifies the base image which
# will be downloaded when you build your image .
FROM pklehre / niso2020 -lab2 - msc
# Add all the files you need for your submission into the Docker image ,
# e.g. source code , Java bytecode , etc. In this example , we assume your
# program is the Python code in the file abc123 .py. For simplicity , we
# copy the file to the /bin directory in the Docker image . You can add
# multiple files if needed .
ADD abc123 .py /bin
# Install all the software required to run your code . The Docker image
# is derived from the Debian Linux distribution . You therefore need to
# use the apt -get package manager to install software . You can install
# e.g. java , python , ghc or whatever you need . You can also
# compile your code if needed .
# Note that Java and Python are already installed in the base image .
# RUN apt -get update
# RUN apt -get -y install python - numpy
# The final line specifies your username and how to start your program .
# Replace abc123 with your real username and python /bin/ abc123 .py
# with what is required to start your program .
CMD ["- username ", " abc123 ", "- submission ", " python /bin / abc123 .py "]
4. Build the Docker image as shown below. The base image pklehre/niso2019-lab3 will be downloaded from Docker Hub
docker build . -t niso2 - abc123
5. Run the docker image to test that your program starts. A battery of test cases will be executed to check your solution.
docker run niso2 - abc123
6. Once you are happy with your solution, compress the directory containing the Dockerfile as a zip-file. The directory should contain the source code, the Dockerfile, and the PDF file for
Exercise 4 and 5. The name of the zip-file should be niso2-abc123.zip (again, replace the
abc123 with your username).
Following the example above, the directory structure contained in the zip file should be as
follows:
niso2-abc123/exercise.pdf
niso2-abc123/abc123.py
niso2-abc123/Dockerfile
Submissions which do not adhere to this directory structure will be rejected!
7. Submit the zip file niso2-abc123.zip on Canvas.