Algorithmic Compositional Techniques Developed for "Pleonid"

Posted on October 30, 2016 by Stuart Popejoy


“Pleonid” is an algorithmic composition for quintet premiered in May 2012. It is generated, arranged and scored entirely by software. This paper describes the largely generative methods employed, including a novel re-working of the “Tone Clock” by Peter Schat, as well as a new use of braids as both a motif- and counterpoint-generating techique.


Pleonid is both a piece that was performed live in May 2012 by musicians reading from a score, as well as a software program. The software builds a piece from three types of sections: “melody” passages instruct musicians to freely interpret written motives; “vamp” sections feature a soloist improvising freely over a fixed, repeating chordal accompaniment; “braid” sections are through-composed, contrapuntal vignettes.

The quintet piece from May 2012 is but a single parameterization of the software, however. The melodic material, “gamut” (a term I use to describe pseudo-octave intervals other than 12-note), the number of instruments, and instrument-transposition scheme (related to gamut) are all examples of parameters that can be changed, producing entirely new if related compositions.

This paper uses the term “Pleonid” to refer to both the software and the piece. When a particular phrase or note configuration is mentioned, this reflects examples drawn from the particular piece. Techniques and methods however refer to the software and any piece that might be generated from it. Appendix A lists all the parameters required by the program.

Generative Techniques for Pitch Sequences

The software’s first task is to generate a body of pitch-sequence material. It uses a single pitch sequence as input to get started.

Seed melody of Pleonid.

Seed melody of Pleonid.

The pitch-class set of the source melody is determined. In Pleonid this is [0,2,3,4,5,7], Forte no. 6-8. However this is not used for pitch-set transformations but as a concrete scale.

Pitch class set and main scale.

Pitch class set and main scale.

Finding all scales in the source scale

The first elaboration finds every “sub-scale” or chord that can be built from the main scale. This is done by computing every ordered sum possible in the scale’s interval vector. For example, the sequence C D E F has intervals [2,2,1]. The resulting tuples are [[2,2,1],[2,3],[4,1],[5]].

Generating all chords in scale C-D-E-F

Generating all chords in scale C-D-E-F

Performing this operation on the Pleonid interval vector [2,1,1,1,2] produces 15 tuples (the final dyad is discarded, as too trivial a “scale”).

Pleonid scale tuples.

Pleonid scale tuples.

Steering Scales: a modified Tone Clock

The scales produced are now used to generate pitch sequences, using techniques inspired by the “Tone Clock” of Dutch composer Peter Schat.1 In the Tone Clock, triads are classified with their inversion: the major triad, represented by the intervals [4,3], forms a class or “hour” with its inverse, the minor triad [3,4], and so on. The 12-tone scale has 12 triad classes or “hours”, thus the Clock metaphor.

Schat then shows that for each triad class, there is a fixed number of configurations by which placements of the triad and its inversion can saturate all 12 notes, or form a row in dodecaphonic terminology. Schat calls these configurations steerings, as though the triad class is being “steered” through the 12-tone space.

For example, the major/minor triad class has exactly two possible steerings: major at 0, minor at 2, major at 6, and minor at 8; and minor at 0, minor at 2, major at 4, major at 6. It is noteworthy that each triad steering produces a tetrad of the triad roots, here [0,2,6,8] and [0,2,4,6].2

Steerings of the major/minor triad class hour in the Tone Clock.

Steerings of the major/minor triad class “hour” in the Tone Clock.

The 12 triad classes produce 33 steerings in total.

Steerings in Pleonid

In Pleonid, Schat’s triad classes are expanded to accept any chord size or “N-ad”. With tetrad and larger chords, inversion is too limited as a classifying method; we use the more general rotation of chord intervals to find all class members. For triads, rotation and inversion are isomorphic: the major triad class ([3,4],[4,3]) represents all rotations of [4,3]. For tetrads, three such rotations are possible: the tetrad [1,3,2] produces the class ([1,3,2],[3,2,1],[2,1,3]). Pentads will have four class members, and so forth.

Rotations of the tetrad [1,3,2].

Rotations of the tetrad [1,3,2].

Additionally, steering is expanded to configure an arbitrary gamut, instead of just the 12-tone space; in Pleonid the gamut of 10 is used. This however means that the full saturation is not always possible, as the chord size might not evenly divide into the gamut: a triad for instance will not “place” into a 10-gamut using up all pitches. Thus, the constraint of full saturation is removed: maximal saturation of the gamut is found, when total saturation is not possible.

Lastly, a steering means there must be at least two placements: if a chord cannot be placed with its class in any way in a gamut, it is rejected altogether.

Steering algorithm

The search for all valid configurations of a tuple class in a gamut begins with the observation that every N-ad is steered by another “M-ad” where M is N integer-divided by the gamut. Thus a triad is steered by a tetrad in the 12-gamut: 12 / 3 = 4. In a 10-gamut, a triad would be steered by some other triad: 10 / 3 = 3 (truncated), and a pentad by a dyad: 10 / 5 = 2.

The search thus enumerates every M-ad in the gamut, and makes candidate steerings by placing each possible N-ad rotation at each position specified by the M-ad. Candidates are filtered for collisions and duplicates, producing the final result.

One interesting result of using gamuts other than 12 is “gamut-wrapping”: chord placements that “wrap” around the gamut boundary are modulo-remaindered to the appropriate gamut degree. In a 12-gamut, this results in the same chroma, but in other gamuts, different chromas result, as illustrated below.

For example, finding all steerings of the 5-tuple [2,1,2,2] in a 10-gamut means searching every possible dyad (90 total) with every possible rotation (4 total) resulting in 360 candidates. Some optimizations are possible, for instance dyads having intervals larger than 1/2 the gamut size are “un-steerable”. For this 5-tuple, just 2 steerings are found.

Steering [2,1,2,2] in a 10-gamut.

Steering [2,1,2,2] in a 10-gamut.

The last tuple of the second steering illustrates gamut-wrapping. The rotation of the interval [2,2,1,2] puts the last B above the gamut boundary of Bb; modulo wrapping results in a C#. This illustrates how non-standard gamuts create melodic interest: in a 12-gamut, wrapping notes above the gamut result in the same chroma; here, having a C# instead of a B adds new pitch information to the composition.

Of Pleonid’s 15 tuples produced above, 3 tuples are un-steerable but the rest are quite productive, producing 49 distinct steerings.

Preserving seed melody features

While all 49 steerings are created from a single scale, the steerings introduce a great deal more pitch information which obscures any relationship to the original material. The attempt to recover and amplify this relationship starts by filtering out sequences that do not have at least one steering chord composed entirely of seed-scale pitches.

From the 5-tuple example above, the second steering ([0,2,3,5,7],[4,6,8,9,1]) is allowed, since the first tuple is made from intervals contained in the [0,2,3,4,5,7] seed scale. However, the first steering ([0,1,3,5,7],[2,4,6,8,9]) is discarded, since both tuples have intervals outside of the seed scale.

This filter reduces the 49 steerings to 32. The tuples then undergo a mapping procedure, designed to map any scale tones back to the register they appear in the original seed sequence. So for instance, if a G appears in a chord, it would be mapped to below middle-C.

Mapping pitches to their register in the original seed sequence.

Mapping pitches to their register in the original seed sequence.

Unfortunately, I did not verify these original results carefully, as I discovered later: a bug in the code in fact results in a fairly chaotic remapping of pitches. 3

Actual mapping in Pleonid code.

Actual mapping in Pleonid code.

Thus, this attempt to preserve seed melody features fails: by mis-assigning pitches found in the original pitch-set, values reflecting the seed melody are more distorted than other values. Nonetheless, it creates more intervallic motion than would otherwise be present.4

Pitch sequences from steerings

The 32 modified steerings are made into pitch sequences or “lines” by arpeggiating the tuples constituting the steering, and then interleaving the results.

Interleaving steerings to generate lines.

Interleaving steerings to generate lines.

Monotonically increasing or decreasing lines are filtered, removing one result. 31 sequences remain, representing the totality of intervallic material used in the composition.

Rhythm + Pitch = Motif: Braids for Rhythmic Generation

The 31 sequences are given rhythmic character to form the main motivic “lines” of the piece. This is accomplished by projecting the pitch sequences onto a strand of a braid.

Braids are 2-dimensional representations of knots, used in a branch of mathematics that seeks to describe and understand the essential features of knots, with applications in bioinformatics, fluid mechanics and cryptography. Braids themselves form a group, but most significantly for Pleonid, they present a cartesian-style left-to-right “plot” of information that resembles music notation or “piano rolls”.

Musical interpretation of a braid strand

A braid is built from “strands” which proceed from left to right, crossing over or under other strands as it weaves through the braid representation. The musical interpretation reads each strand as a pitch sequence changing in time, with the x-axis representing a time quanta (an eighth-note in Pleonid), and the y-axis indexing pitch degrees from some pitchset.

A braid from Pleonid. A column or step represents an eighth-note duration, with vertical positions representing specific pitches.

A braid from Pleonid. A column or “step” represents an eighth-note duration, with vertical positions representing specific pitches.

The polarity of a cross – whether a strand passes over or under another strand – is used to decide when a strand sounds a particular pitch. An over-cross or flat cross sounds the indexed pitch, whereas an under-cross continues the pitch started in a previous cross. Under-crosses at the beginning of a strand are removed.5 For instance, in the braid above, the strand starting at 10 begins with an over cross, sounding the pitch 10 through the following 4 under-crosses. It then over-crosses at 1, and again at 5, resulting in two new notes.

Projecting a pitch sequence as a strand

Using the rules above, the generated pitch sequences are each “projected” as a strand of a braid. The sequence is analyzed to generate a pitch index for the Y-axis, and the strand starts at the first pitch, immediately travelling to the second. This method results in pitch transitions taking longer for more “remote” pitches.

The source line.

The source line.

The strand generated from the source line.

The strand generated from the source line.

The resulting melody.

The resulting melody.

The last note of a strand presents a problem with this interpretation, as it has no length information. In Pleonid, we simply assign this last note the minimum duration, an eighth-note.

By projecting the 31 sequences onto strands, we have now provided rhythmic information to arrive at 31 melodies or motives. These are used in the “interpretive” sections of the piece, where a musician uses a melody to ground structured improvisation.

Elaboration: Braids for Counterpoint

The creation of a strand actually specifies an entire braid in itself, but not a very musically-interesting one, with other strands sustaining their indexed pitch until they encounter the source strand.

The strand above as a minimal braid.

The strand above as a minimal braid.

In Pleonid, braids are further elaborated using a “terracing” strategy, vertically inverting and reflecting the source strand material.

Each transition in the source strand is seen as a “column” or “step”, being either UP, DOWN or FLAT; the cross type is identified as OVER or UNDER. Moving away vertically from the strand step, FLATs and crosses are alternated; when a cross is generated, it is “reflected” from the previous cross – a DOWN becomes an UP and an OVER becomes an UNDER. The alternating flats and inverting crosses produces a visually “terraced” effect. At the vertical edges, duplicate FLATs are inserted if a cross would violate the braid boundary. 6

Elaborating a braid. The terracing result is shown on every-other step to illustrate the columnar approach.

Elaborating a braid. The “terracing” result is shown on every-other step to illustrate the columnar approach.

The resulting braid.

The resulting braid.

The resulting braid now has more motion in the non-source strands, and importantly has more looping than the minimal braid.

Braid Loops

When a braid strand terminates at a different Y-position than where it started, it is seen as joining the strand starting at the ending Y-position of the first. The integrity of the braid representation means that this strand will too end at a different position, eventually forming a “loop” that joins back with the starting strand. A complex braid of many strands may indeed represent a single loop.

A braid with 7 strands, but only 3 loops, differentiated by color.

A braid with 7 strands, but only 3 loops, differentiated by color.

Loops have compelling musical properties, as interpreted in Pleonid: a loop length will always be a multiple of the “braid length” or number of steps. If the braid length is a “bar”, single-strand loops repeat every bar, while 2-loops only repeat every two bars, and so forth, creating easily discernable temporal interplay between parts. In Pleonid loops, not strands, are assigned to parts, so two braids of the same “height” will often have different polyphony.

Braid loops realized. The first staff is a 4-bar phrase, the second 1-bar, the third 2-bar.

Braid loops realized. The first staff is a 4-bar phrase, the second 1-bar, the third 2-bar.

In the figure, the braid loops shown above are orchestrated for three parts, using a 10-semitone register divide. The top staff has the longest loop, lasting the full 4 bars. The second staff has the shortest loop, repeating a one-bar figure 4 times. The third staff has a 2-bar figure, repeated twice.

The 31 melodic figures thus create 31 contrapuntal passages, forming the “braid” sections of the piece.

Braid “Vamps”

The final generative procedure filters each of the 31 braid passages to find simultaneous note onsets across all parts. The note lengths are matched to the longest note in the onset group.

A braid realization with simultaenous note onsets colored in red.

A braid realization with simultaenous note onsets colored in red.

The resulting chordal passage. For each chord, the longest note found determines the ensemble length.

The resulting chordal passage. For each chord, the longest note found determines the ensemble length.

These form the “vamp” sections. As the chord sections are generated, the software keeps track of a “soloist” index which is advanced for each section, skipping those who are playing; this selects an instrument to be instructed to improvise freely over the chordal accompaniment.

Arrangement and Orchestration

With all musical material now generated, the piece is ready to be assembled.

Voicing instruments

As each section is generated, it is “voiced” for the quintet by offsetting each part by the gamut “register” of 10 semitones.

Pleonid is arranged for violin, alto saxophone, vibraphone, trombone and electric bass, with the generated material being generated around C5, in the soprano range. Thus as the part index increases, 10 semitones are subtracted, which worked well for the instruments except vibraphone. Some notes fell beneath the vibraphone range, for which special-purpose code adjusts individual bars up as necessary.

As seen in the “chordal” sections, the gamut-register 10 results in sonorous, “wide” voicings with plenty of intervallic interest.

Pseudorandom behavior with a LFSR

The construction of the piece is by pseudo-random selection of one of three section types, representing the three bodies of material generated: lines, chords, and braids. Preferring a parameterizable, deterministic and low-entropy solution, I use a linear feedback shift register (LFSR) to produce a stream of on-off values. In Pleonid, this LFSR is parameterized with a bit length of 5 and tap points of 2 and 4.

data LFSR = LFSR { _lState :: [Bool],
                   _lTap1 :: Int,
                   _lTap2 :: Int }
            deriving (Eq,Show)

-- | Create an LFSR with initial True values.
-- | Errors not handled (like taps > length) for brevity.
initLFSR :: Int -> Int -> Int -> LFSR
initLFSR len tap1 tap2 = LFSR (replicate len True) tap1 tap2

-- | Feed head with XOR of tap points.
shift :: LFSR -> LFSR
shift (LFSR register tap1 tap2) = LFSR (gen:init register) tap1 tap2
    where gen = register !! tap1 /= register !! tap2

> mapM_ (print . _lState) $ take 6 $ iterate shift $ initLFSR 5 2 4

The code example shows the Pleonid LFSR stepping through six iterations, pushing a new boolean value atop the 5-value list each time.

Stepwise section selection

Section type selection is controlled by two values, the section type itself, which can be in the ordered range of [Chord,Line,Braid], and a “step direction”, an integer value that ranges between (-1,0,1). The LFSR is sampled, which then increments the step direction up or down, wrapping around such that a False value with a direction of -1 wraps to 1, and so forth. This direction is then applied to the section type value, which in turn wraps: Braid increments up to Chord and down to Line. The effect is to “slew” or smooth the random LFSR to follow the section type sequence in one direction or another.

In Pleonid, the section type is initialized to Chord and the direction to 0. The LFSR is iterated and sampled, which gives False. This results in a direction of -1, which wraps the section type to Braid. Thus the piece starts with a Braid section. The first 5 bars are summarized in the following table:

LFSR Direction Type
False -1 Braid
False 1 (wrapped) Chord (wrapped)
False 0 Chord
True 1 Line
True -1 (wrapped) Chord

Section rendering

The “lines” sections are the only ones requiring assembly at this late stage. The LFSR’s length comes into play here, as it is used to select which instruments will play in a given lines section. When the fourth section is dertermined to be Line, the LFSR reads [True,False,False,False,True]. This means the violin (instrument 1) and the bass (instrument 5) are selected to play together. The first two lines are “popped off” the lines list and given to the two instruments to interpret. As a result, lines are used much more frequently than chord or braid sections, so the lines are stored in a “cycling list”, repeating indefinitely.

Meanwhile, when a chord or braid section is selected, that section is “popped off” the list permanently, and the piece ends when one of these lists is empty. in Pleonid, the braids run out first (with only 2 chord sections left), so the piece ends with a braid section. In total, 79 sections are thus generated.

Scoring and representation

Scoring is accomplished by rendering the music data as MusicXML, and using Sibelius software to load and print the parts. MusicXML representation can be done with my open-source libraries which are listed in Appendix B.

The hard part of output is not the XML transformation, but beating the data into a shape that can be read by musicians. Toward this end, a lot of work is done to have proper note-lengths (ie, a 5/8 note is not notatable, but must be written as a quarter tied to an eighth) and to auto-select bar lengths to avoid brutal time signatures like 21/8. This work is beyond the scope of this paper, but code for tie rules and re-barring phrases can be found in the open source code.


Pleonid represents my most far-reaching work in algorithmic composition, but in truth I’ve only scratched the surface. Just exploring the possibilities of the parameters alone represents new works waiting to be realized.

Braids also have far more promise than just my treatment, as my musical interpretation is by no means the only way to use this way to structure data with such formal regularity; work can be done also on using some of the braid isotopy transformations (Reidemeister moves) to see what kinds of contrapuntal transformations might ensue.

The purpose of this paper is to document these approaches in hopes that they will inspire new work by more composers in the realm of generative, algorithmic software. I also feel like the interaction of algorithmic composition with improvisational performance is very fertile, as reflected also in the work of Anthony Braxton and others.

Appendix A: Input Parameters

The Pleonid software accepts the following input parameters:

Parameter Description
gamut Gamut pseudo-octave value, used in steerings.
line Source pitch sequence.
transposer Voicing/register function, in Pleonid set to offset by 10 semitones.
scaleMapper Function to map steerings to scale pitches, parameterized to mimic bug in the original Java code or use new “fixed” version.
filterRestsInBraids When true, drop any initial “under” crosses, resulting in rhythmic disparities in the original codebase. False instead puts a rest in its place.
scaleSize A separate gamut value used in scale mapping, and in pitch-set discovery. Pleonid uses the standard 12-octave.
lfsr A parameterized LFSR object. Pleonid’s is parameterized with 5 slots with taps at 2 and 4.
parts A list of part metadata, including names and clefs, also dictating the number of parts.

Appendix B: Relevant open-source libraries by the author


Provides music primitives, like a polymorphic note and note-list “bar” datatype. Uses fadno-xml to render output to MusicXML, and can also produce midi files. Also supports the advanced barring and note-tying features used in Pleonid in the Fadno.Meter module.


Provides the braid representations used by the Haskell version of Pleonid, plus graphics utilities used for this paper.


Provides a type-model of MusicXML 2.0, generated automatically from the music-xml-2.0 xsd. Attempts to provide a general-purpose XSD-to-datamodel code generator, but that so far only works on MusicXML :)

  1. Schat, Peter (1993). Also see and Thanks to Jochem van Dijk for hipping me to it.

  2. For Schat the tetrads produced by a steering can be related intervallically to other triad hours in the clock, producing a relationship between hours.

  3. The bug was discovered only in a later effort to migrate the entire Pleonid program from Java to Haskell. While I’m tempted to credit Haskell’s superior expressiveness as the reason for discovering the bug, it’s really because of the extensive code-review that resulted from performing the migration.

  4. The Haskell code base in fact preserves all of the original bugs and quirks of the Java codebase, in order to exactly reproduce the Pleonid piece; however as the appendix shows, it is possible to parameterize the new codebase to operate “bug-free”, producing a different piece.

  5. Handling of initial under-crosses is perhaps a bug, as by removing them, the rhythmic projection is distorted. A parameter in the Haskell code now governs whether to remove or convert to an initial rest.

  6. Here too a “bug” was discovered in the original code, where the terracing procedure was slightly different when moving downward from the strand as opposed to upward.