Genetic
Programming
System for Music
Generation
With Automated Fitness Raters
What is Genetic
Programming(GP)?
• an evolutionary algorithm-based methodology
inspired by biological evolution
• New generation of solutions created from previous
generations
• Three Steps:
o Mutation: change random bits
o Crossover: exchange parts of two parents
o Selection: get rid of bad examples
An Example: 8-Queen
Problem
 Fitness Function:
 How good is each string?
 In this case, it depends on the number of non-attacking pairs of
queens.
 Selection:
 Choose parents randomly according to fitness function
 Crossover:
 Make new children from parents
 Choose a cut-point, swap halves
 Mutation
 Change random bits with low probability
Back to the GP-Music
System
 An interactive system which allows users to evolve
short musical sequences.
 Focus on creating short melodic sequences.
 Does not attempt to evolve polyphony or the
actual wave forms of the instruments.
 Create only a set of notes and pauses.
 Use XM format - store musical pieces rather than
straight digital audio
Function and Terminal
Sets
• Function Set:
play_two(2 arguments),
add_space(1 argument), play_twice(1 argument),
shift_up(1 argument), shift_down(1 argument),
mirror(1 argument), play_and_mirror(1 argument)
• Terminal Set:
o Notes: C-4, C#4, D-4, D#4, E-4, F-4, F#4, G-4, G#4, A-5, A#5, B-5
• Pseudo-Chords:
o C-Chord, D-Chord, E-Chord, F-Chord, G-Chord, A-Chord, B-Chord
• Other: RST (used to indicates one beat without a
note)
Function and Terminal
Sets Cont.
Pseudo-Chord
Corresponding Note Sequence
C-Chord
C-4, E-4, G-4
D-Chord
D-4, F#4, A-5
E-Chord
E-4, G#4, B-5
F-Chord
F-4, A-5, C-5
G-Chord
G-4, B-5, D-5
A-Chord
A-5, C#5, E-5
B-Chord
B-5, D#5, F#5
Sample Music Program
Interpreter:
(shift-down (add-space (play-and-mirror (play-two (play-two (play-two (play-two
B-5 B-5) (shift-down A-5)) (shift-down A-5)) F-4))))
A-5,RST,A-5,RST,F-4,RST,F-4,RST,E4,RST,
E-4,RST,F-4,RST,F-4,RST,A-5,RST,A5,RST
B-5,RST,B-5,RST,G-4,RST,G-4,RST,F4,RST,
F-4,RST,G-4,RST,G-4,RST,B-5,RST,B5,RST
B-5,B-5,G-4,G-4,F-4,F-4,G-4,G-4,B-5,B-5
shift-down
add-space
play-andmirror
B-5,B-5,G-4,G-4,F-4
play-two
B-5,B-5,G-4,G-4
play-two
B-5,B-5,G-4
B-5,B-5
play-two
play-two
B-5
B-5
F-4
shift-down
shift-down
A-5
A-5
Each node in the tree propagates
up a musical note string, which
is then modified by the next
higher node. In this way a
complete sequence of notes is
built up, and the final string is
returned by the root node.
Fitness Selection
A human using the system is asked to rate the musical sequences
that are created for each generation of the GP process.
User Bottleneck
 A user can only rate a small number of sequences
in a sitting, limiting the number of individuals and
generations that can be used.
 The GP-Music System takes rating data from a users
run and uses it to train a neural network based
automatic rater.
Neural Networks
• A trainable mathematical model for finding
boundaries
• Inspired by biological neurons
o Neurons collect input from receptors, other neurons
o If enough stimulus collected, the neuron fires
• Inputs in artificial neurons are variables, outputs of
other neurons
• Output is an “activation” determined by the
weighted inputs
Back Propagation
Algorithm
 We have some error and we want to assign
“blame” for it proportionally to the various weights
in the network
 We can compute the error derivative for the last
layer
 Then, distribute “blame” to previous layer
Backprop Training
Start with random weights
Run the network forward
Calculate the error based on outputs
Propagate the error backward
Update the weights
Repeat with next training example until you stop
improving
o Cross validation data is very important here
o
o
o
o
o
o
each of the top level nodes has ‘Level
Spread’ connections to lower nodes.
The weights on these connections are
all shared, so the weight on the first
input to each upper level node is
identical
each lower level node affects
two upper level nodes
Auto Rater Runs
• The weights and biases for the auto-rater network
trained for 850 cycles were used in several runs of
the GP-Music System.
• To evaluate how well the auto-rater works in larger
runs, runs with 100 and 500 sequences per
generation over 50 generations were made. The
resulting best individuals are shown in the next slide
50 Generations, 100 individuals per Generation
50 Generations, 500 individuals per
Generation
Future Work
• it would be interesting to analyze the weights which
are being learned by the network to see what sort
of features it is looking for
• It may also be possible to improve the structure of
the auto raters themselves by feeding them extra
information, or modifying their topology
Reference
The research is done by
Brad Johanson - Stanford University
Rains Apt. 9A
704 Campus Dr.
Stanford, CA. 94305
bjohanso@stanford.edu
Riccardo Poli - University of Birmingham
School of Computer Science
The University of Birmingham
Birmingham B15 2TT
R.Poli@cs.bham.ac.uk