Computational Creativity - IIIA CSIC

Share

Transcript

1 ARBOR Ciencia, Pensamiento y Cultura Vol. 189-764, noviembre-diciembre 2013, a082 | ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005 EL LEGADO DE ALAN TURING / THE LEGACY OF ALAN TURING COMPUTATIONAL CREATIVITY CREATIVIDAD COMPUTACIONAL Ramon Lpez de Mntaras Badia Instituto de Investigacin en Inteligencia Artificial Consejo Superior de Investigaciones Cientficas Campus UAB, 08193 Bellaterra E-mail: [email protected] Citation/Cmo citar este artculo: Lpez de Mntaras Badia, Copyright: 2013 CSIC. This is an open-access article distribu- R. (2013). Computational Creativity. Arbor, 189 (764): a082. ted under the terms of the Creative Commons Attribution-Non doi: http://dx.doi.org/10.3989/arbor.2013.764n6005 Commercial (by-nc) Spain 3.0 License. Received: 10 July 2013. Accepted: 15 September 2013. ABSTRACT: New technologies, and in particular artificial intel- RESUMEN: Las nuevas tecnologas y en particular la Inteligencia ligence, are drastically changing the nature of creative proc- Artificial estn cambiando de forma importante la naturaleza esses. Computers are playing very significant roles in creative del proceso creativo. Los ordenadores estn jugando un papel activities such as music, architecture, fine arts, and science. muy significativo en actividades artsticas tales como la msica, Indeed, the computer is already a canvas, a brush, a musical la arquitectura, las bellas artes y la ciencia. Efectivamente, el instrument, and so on. However, we believe that we must aim ordenador ya es el lienzo, el pincel, el instrumento musical, etc. at more ambitious relations between computers and creativity. Sin embargo creemos que debemos aspirar a relaciones ms Rather than just seeing the computer as a tool to help human ambiciosas entre los ordenadores y la creatividad. En lugar de creators, we could see it as a creative entity in its own right. This verlos solamente como herramientas de ayuda a la creacin, view has triggered a new subfield of Artificial Intelligence called los ordenadores podran ser considerados agentes creativos. Computational Creativity. This article addresses the question Este punto de vista ha dado lugar a un nuevo subcampo de la of the possibility of achieving computational creativity through Inteligencia Artificial denominado Creatividad Computacional. some examples of computer programs capable of replicating En este artculo abordamos la cuestin de la posibilidad de de some aspects of creative behavior in the fields of music and alcanzar dicha creatividad computacional mediante algunos science. ejemplos de programas de ordenador capaces de replicar algu- nos aspectos relacionados con el comportamiento creativo en los mbitos de la msica y la ciencia. KEYWORDS: Artificial Intelligence, Computational Creativity. PALABRAS CLAVE: Inteligencia Artificial, Creatividad Compu- tacional.

2 INTRODUCTION process, and not just the outcome of it, is taken into account when assessing artworks. Hence, one could Computational creativity is the study of building argue that such Turing-style tests are essentially set- a082 software that exhibits behaviour that would be dee- ting the Computers up for a fall. med creative in humans. Such creative software can be used for autonomous creative tasks, such as inven- Creativity seems mysterious because when we have Computational Creativity ting mathematical theories, writing poems, painting creative ideas it is very difficult to explain how we pictures, and composing music. However, computa- got them and we often talk about vague notions like tional creativity studies also enable us to understand inspiration and intuition when we try to explain human creativity and to produce programs for creati- creativity. The fact that we are not conscious of how ve people to use, where the software acts as a creative a creative idea manifests itself does not necessarily collaborator rather than a mere tool. Historically, its imply that a scientific explanation cannot exist. As a been difficult for society to come to terms with machi- matter of fact, we are not aware of how we perform nes that purport to be intelligent and even more diffi- other activities such as language understanding, pat- cult to admit that they might be creative. Even within tern recognition, and so on, but we have better and computer science, people are still sceptical about the better AI techniques able to replicate such activities. creative potential of software. A typical statement of Since nothing can arise from the emptiness, we detractors of computational creativity is that simula- must understand that every creative work or creative ting artistic techniques means also simulating human idea always is preceded by a historical - cultural sche- thinking and reasoning, especially creative thinking. me, it is a fruit of the cultural inheritance and the lived This is impossible to do using algorithms or informa- experiences. As Margaret Boden states in her book Ar- tion processing systems. We couldnt disagree more. tificial Intelligence and Natural Man: As is hopefully evident from the examples in this pa- per, creativity is not some mystical gift that is beyond Probably the new thoughts that originate in the mind scientific study but rather something that can be in- are not completely new, because have their seeds in vestigated, simulated, and harnessed for the good of representations that already are in the mind. To put it differently, the germ of our culture, all our knowled- society. And while society might still be catching up, ge and our experience, is behind each creative idea. computational creativity as a discipline has come of The greater the knowledge and the experience, the age. This maturity is evident in the amount of activi- greater the possibility of finding an unthinkable rela- ty related to computational creativity in recent years; tion that leads to a creative idea. If we understand in the sophistication of the creative software we are creativity like the result of establishing new relations building; in the cultural value of the artefacts being between pieces of knowledge that we already have, produced by our software; and most importantly, in then the more previous knowledge one has the more the consensus we are finding on general issues of capacity to be creative. computational creativity. With this understanding in mind, an operational, Computational creativity is a very lively subject and widely accepted, definition of creativity is: a area, with many issues still open to debate. For ins- creative idea is a novel and valuable combination of tance, many people still turn to the Turing test (Tu- known ideas. In other words, physical laws, theo- ring, 1950) to approximate the value of the artefacts rems, musical pieces can be generated from a finite produced by their software. That is, if a certain num- set of existing elements and, therefore, creativity is an ber of people cannot determine which artefacts were advanced form of problem solving that involves me- produced by computer and which were produced by a mory, analogy, learning, and reasoning under constra- human, then the software is doing well. Other people ints, among other things, and is therefore possible to believe that the Turing test is inappropriate for creati- replicate by means of computers. ve software. One has to ask the question, Under full This article addresses the question of the possibility disclosure, would people value the artefacts produ- of achieving computational creativity through some ced by a computer as highly as they would the human examples of computer programs capable of replica- produced ones? In some domains, the answer could ting some aspects of creative behaviour in the fields be yes: for instance, a joke is still funny whether or not of music and science. We did not intend to cover the it is produced by a computer. In other domains, such full range of AI approaches to computational creativi- as the visual arts, however, the answer is very likely ty and we could not include the many existing areas to be no. This highlights the fact that the production of application. 2 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

3 For further reading regarding computational crea- Moorers program generated simple melodies, along tivity in general, I recommend the books by Boden with the underlying harmonic progressions, with sim- (1991, 1994, 2009), Dartnall (1994), Partridge and ple internal repetition patterns of notes. This appro- a082 Rowe (1994), and Bentley and Corne (2002); as well ach relies on simulating human composition processes as the papers by Rowe and Partridge (1993), Bucha- using heuristic techniques rather than on Markovian Ramon Lpez de Mntaras Badia nan (2001) and the recent special issue of AI Maga- probability chains. Levitt (1983) also avoided the use zine edited by Colton, Lpez de Mntaras and Stock of probabilities in the composition process. He argues (2009). that: randomness tends to obscure rather than reveal the musical constraints needed to represent simple COMPUTATIONAL CREATIVITY IN MUSIC musical structures. His work is based on constraint- based descriptions of musical styles. He developed a Artificial Intelligence has played a crucial role in description language that allows musically meaningful the history of computer music almost since its begin- transformations of inputs, such as chord progressions nings in the 1950s. However, until quite recently, most and melodic lines, to be expressed through a series of effort had been on compositional and improvisational constraint relationships that he calls style templates. systems and little efforts had been devoted to expres- He applied this approach to describe a traditional jazz sive performance. In this section we review a selec- walking bass player simulation as well as a two-handed tion of some significant achievements in AI appro- ragtime piano simulation. aches to music composition, music performance, and improvisation, with an emphasis on the synthesis of The early systems by Hiller-Isaacson and Moorer expressive music. were both based also on heuristic approaches. Howe- ver, possibly the most genuine example of early use Composing music of AI techniques is the work of Rader (1974). Rader used rule-based AI programming in his musical round Hiller and Isaacson (1958) work, on the ILLIAC com- (a circle canon such as Frre Jacques) generator. The puter, is the best known pioneering work in compu- generation of the melody and the harmony were ba- ter music. Their chief result is the Illiac Suite, a string sed on rules describing how notes or chords may be quartet composed following the generate and test put together. The most interesting AI component of problem solving approach. The program generated this system are the applicability rules, determining notes pseudo-randomly by means of Markov chains. the applicability of the melody and chord generation The generated notes were next tested by means of rules, and the weighting rules indicating the likeliho- heuristic compositional rules of classical harmony od of application of an applicable rule by means of a and counterpoint. Only the notes satisfying the rules weight. We can already appreciate the use of metak- were kept. If none of the generated notes satisfied nowledge in this early work. the rules, a simple backtracking procedure was used to erase the entire composition up to that point, and AI pioneers such as Herbert Simon or Marvin Mins- a new cycle was started again. The goals of Hiller and ky also published works relevant to computer music. Isaacson excluded anything related to expressiveness Simon and Sumner (1968) describe a formal pattern and emotional content. In an interview (Schwanauer language for music, as well as a pattern induction and Levitt, 1993, p. 21), Hiller and Isaacson said that, method, to discover patterns more or less implicit before addressing the expressiveness issue, simpler in musical works. One example of pattern that can problems needed to be handled first. We believe that be discovered is the opening section is in C Major, this was a very correct observation in the fifties. Af- it is followed by a section in dominant and then a re- ter this seminal work, many other researchers based turn to the original key. Although the program was their computer compositions on Markov probability not completed, it is worth noticing that it was one of transitions but also with rather limited success jud- the firsts in dealing with the important issue of music ging from the standpoint of melodic quality. Indeed, modelling, a subject that has been, and still is, widely methods relying too heavily on markovian processes studied. For example, the use of models based on ge- are not informed enough to produce high quality mu- nerative grammars has been, and continues to be, an sic consistently. important and very useful approach in music mode- lling (Lerdahl and Jackendoff, 1983). However, not all the early work on composition re- lies on probabilistic approaches. A good example is the Marvin Minsky, in his well known paper Music, work of Moorer (1972) on tonal melody generation. Mind, and Meaning (Minsky, 1981), addresses the 3 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

4 important question of how music impresses our In HARMONET (Feulner, 1993), the harmonization minds. He applies his concepts of agent and its role in problem is approached using a combination of neural a society of agents as a possible approach to shed light networks and constraint satisfaction techniques. The a082 on that question. For example, he hints that one agent neural network learns what is known as the harmo- might do nothing more than noticing that the music nic functionality of the chords (chords can play the Computational Creativity has a particular rhythm. Other agents might perceive function of tonic, dominant, subdominant, etc) and small musical patterns such as repetitions of a pitch; constraints are used to fill the inner voices of the differences such as the same sequence of notes pla- chords. The work on HARMONET was extended in yed one fifth higher, etc. His approach also accounts the MELONET system (Hrnel and Degenhardt, 1997; for more complex relations within a musical piece by Hrnel and Menzael, 1998). MELONET uses a neural means of higher order agents capable of recognizing network to learn and reproduce higher-level structure large sections of music. It is important to clarify that in in melodic sequences. Given a melody, the system in- that paper Minsky does not try to convince the reader vents a baroque-style harmonization and variation of about the question of the validity of his approach, he any chorale voice. According to the authors, HARMO- just hints at its plausibility. NET and MELONET together form a powerful music- composition system that generates variations whose Among the compositional systems there is a large quality is similar to those of an experienced human number dealing with the problem of automatic har- organist. monization using several AI techniques. One of the earliest works is that of Rothgeb (1969). He wrote a Pachet and Roy (1998) also used constraint satis- SNOBOL program to solve the problem of harmonizing faction techniques for harmonization. These techni- the unfigured bass (given a sequence of bass notes ques exploit the fact that both the melody and the infer the chords and voice leadings that accompany harmonization knowledge impose constraints on the those bass notes) by means of a set of rules such as If possible chords. Efficiency is, however, a problem with the bass of a triad descends a semitone, then the next purely constraint satisfaction approaches. bass note has a sixth. The main goal of Rothgeb was In (Sabater et al., 1998), the problem of harmo- not the automatic harmonization itself but to test the nization is approached using a combination of rules computational soundness of two bass harmonization and case-based reasoning. This approach is based on theories from the eighteenth century. the observation that purely rule-based harmoniza- One of the most complete works on harmonization tion usually fails because, in general, the rules dont is that of Ebcioglu (1993). He developed an expert sys- make the music, it is the music that makes the rules. tem, CHORAL, to harmonize chorales in the style of Then, instead of relying only on a set of imperfect ru- J.S. Bach. CHORAL is given a melody and produces the les, why not making use of the source of the rules, corresponding harmonization using heuristic rules and that is the compositions themselves? Case-based rea- constraints. The system was implemented using a logic soning allows the use of examples of already harmo- programming language designed by the author. An im- nized compositions as cases for new harmonizations. portant aspect of this work is the use of sets of logical The system harmonizes a given melody by first looking primitives to represent the different viewpoints of the for similar, already harmonized, cases, when this fails, music (chords view, time-slice view, melodic view, etc.). it looks for applicable general rules of harmony. If no This was done to tackle the problem of representing rule is applicable, the system fails and backtracks to large amounts of complex musical knowledge. the previous decision point. The experiments have shown that the combination of rules and cases re- MUSACT (Bharucha, 1993) uses Neural Networks to sults in much fewer failures in finding an appropria- learn a model of musical harmony. It was designed to te harmonization than using either technique alone. capture musical intuitions of harmonic qualities. For Another advantage of the case-based approach is that example, one of the qualities of a dominant chord is each newly correctly harmonized piece can be memo- to create in the listener the expectancy that the tonic rized and made available as a new example to harmo- chord is about to be heard. The greater the expectan- nize other melodies; that is, a learning-by-experience cy, the greater the feeling of consonance of the tonic process takes place. Indeed, the more examples the chord. Composers may choose to satisfy or violate system has, the less often the system needs to resort these expectancies to varying degree. MUSACT is ca- to the rules and therefore it fails less. MUSE (Schwa- pable of learning such qualities and generate graded nauer, 1993) is also a learning system that extends expectancies in a given harmonic context. 4 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

5 an initially small set of voice leading constraints by expert system to determine the tempo and the articu- learning a set of rules of voice doubling and voice lea- lation to be applied when playing Bachs fugues from ding. It learns by reordering the rules agenda and by The Well-Tempered Clavier. The rules were obtained a082 chunking the rules that satisfy the set of voice leading from two expert human performers. The output gi- constraints. MUSE successfully learned some of the ves the base tempo value and a list of performance Ramon Lpez de Mntaras Badia standard rules of voice leading included in traditional instructions on note duration and articulation that books of tonal music. should be followed by a human player. The results very much coincide with the instructions given in well Morales-Manzanares et al. (2001) developed a sys- known commented editions of The Well-Tempered tem called SICIB capable of composing music using Clavier. The main limitation of this system is its lack body movements. This system uses data from sensors of generality because it only works well for fugues attached to the dancer and applies inference rules to written on a 4/4 meter. For different meters, the rules couple the gestures with the music in real time. would be different. Another obvious consequence of Certainly the best-known work on computer com- this lack of generality is that the rules are only appli- position using AI is David Copes EMI project (Cope, cable to Bach fugues. 1987, 1990). This work focuses on the emulation of The work of the KTH group from Stockholm (Friberg, styles of various composers. It has successfully com- 1995; Friberg et al., 1998, 2000; Bresin, 2001) is one posed music in the styles of Cope, Mozart, Palestri- of the best-known long-term efforts on performance na, Albinoni, Brahms, Debussy, Bach, Rachmaninoff, systems. Their current Director Musices system incor- Chopin, Stravinsky, and Bartok. It works by searching porates rules for tempo, dynamic, and articulation for recurrent patterns in several (at least two) wor- transformations constrained to MIDI. These rules are ks of a given composer. The discovered patterns are inferred both from theoretical musical knowledge and called signatures. Since signatures are location de- experimentally by training, specially using the so-ca- pendent, EMI uses one of the composers works as a lled analysis-by-synthesis approach. The rules are divi- guide to fix them to their appropriate locations when ded in three main classes: Differentiation rules, which composing a new piece. To compose the musical mo- enhance the differences between scale tones; Grou- tives between signatures, EMI uses a compositional ping rules, which show what tones belong together; rule analyzer to discover the constraints used by the and Ensemble rules, that synchronize the various voi- composer in his works. This analyzer counts musical ces in an ensemble. events such as voice leading directions; use of repea- ted notes, etc. and represents them as a statistical Canazza et al. (1997) developed a system to analyze model of the analyzed works. The program follows how the musicians expressive intentions are reflected this model to compose the motives to be inserted in in the performance. The analysis reveals two different the empty spaces between signatures. To properly expressive dimensions: one related to the energy insert them, EMI has to deal with problems such as: (dynamics) and the other one related to the kinetics linking initial and concluding parts of the signatures to (rubato) of the piece. The authors also developed a the surrounding motives avoiding stylistic anomalies, program for generating expressive performances ac- maintaining voice motions, maintaining notes within cording to these two dimensions. a range, etc. Proper insertion is achieved by means The work of Dannenberg and Derenyi (1998) is also of an Augmented Transition Network (Woods, 1970). a good example of articulation transformations using The results, although not perfect, are quite consistent manually constructed rules. They developed a trum- with the style of the composer. pet synthesizer that combines a physical model with a performance model. The goal of the performance Synthesizing expressive music model is to generate control information for the phy- One of the main limitations of computer-generated sical model by means of a collection of rules manually music has been its lack of expressiveness, that is, lack extracted from the analysis of a collection of contro- of gesture. Gesture is what musicians call the nuan- lled recordings of human performance. ces of performance that are uniquely and subtly inter- Another approach taken for performing tempo and pretive or, in other words, creative. dynamics transformation is the use of neural network One of the first attempts to address expressiveness techniques. In Bresin (1998), a system that combines in music is that of Johnson (1992). She developed an symbolic decision rules with neural networks is im- 5 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

6 plemented for simulating the style of real piano per- match those of the most similar retrieved note. Each formers. The outputs of the neural networks express note in the case base is annotated with its role in time and loudness deviations. These neural networks the musical phrase it belongs to, as well as with its a082 extend the standard feed-forward network trained expressive values. Furthermore, cases do not contain with the back propagation algorithm with feedback just information on each single note but they include Computational Creativity connections from the output neurons to the input contextual knowledge at the phrase level. Therefore, neurons. cases in this system have a complex object-centered representation. We can see that, except for the work of the KTH group that considers three expressive resources, the Although limited to monophonic performances, other systems are limited to two resources such as the results are very convincing and demonstrate that rubato and dynamics, or rubato and articulation. This CBR is a very powerful methodology to directly use limitation has to do with the use of rules. Indeed, the the knowledge of a human performer that is implicit main problem with the rule-based approaches is that in her playing examples rather than trying to make it is very difficult to find rules general enough to cap- this knowledge explicit by means of rules. Some au- ture the variety present in different performances of dio results can be listened at http://www.iiia.csic. the same piece by the same musician and even the es/%7Earcos/noos/Demos/Example.html. More re- variety within a single performance (Kendall and Car- cent papers (Arcos and Lpez de Mntaras, 2001; L- terette, 1990). Furthermore, the different expressive pez de Mntaras and Arcos, 2002), describe this sys- resources interact with each other. That is, the rules tem in great detail. for dynamics alone change when rubato is also taken Based on the work on SaxEx, we developed Tem- into account. Obviously, due to this interdependency, poExpress (Grachten et al. 2004), a case-based reaso- the more expressive resources one tries to model, the ning system for applying musically acceptable tempo more difficult it is to find the appropriate rules. transformations to monophonic audio recordings of We developed a system called SaxEx (Arcos et al., musical performances. TempoExpress has a rich des- 1998), a computer program capable of synthesizing cription of the musical expressivity of the performan- high quality expressive tenor sax solo performances ces, that includes not only timing deviations of perfor- of jazz ballads based on cases representing human med score notes, but also represents more rigorous solo performances. As mentioned above, previous kinds of expressivity such as note ornamentation, rule-based approaches to that problem could not deal consolidation, and fragmentation. Within the tempo with more than two expressive parameters (such as transformation process, the expressivity of the perfor- dynamics and rubato) because it is too difficult to find mance is adjusted in such a way that the result sounds rules general enough to capture the variety present natural for the new tempo. A case base of previously in expressive performances. Besides, the different ex- performed melodies is used to infer the appropriate pressive parameters interact with each other making expressivity. The problem of changing the tempo of it even more difficult to find appropriate rules taking a musical performance is not as trivial as it may seem into account these interactions. because it involves a lot of musical knowledge and creative thinking. Indeed, when a musician performs With CBR, we have shown that it is possible to deal a musical piece at different tempos the performances with the five most important expressive parameters: are not just time-scaled versions of each other (as if dynamics, rubato, vibrato, articulation, and attack of the same performance were played back at different the notes. To do so, SaxEx uses a case memory con- speeds). Together with the changes of tempo, varia- taining examples of human performances, analyzed tions in musical expression are made (Desain and by means of spectral modeling techniques and back- Honing, 1993). Such variations do not only affect the ground musical knowledge. The score of the piece to timing of the notes, but can also involve for example be performed is also provided to the system. The core the addition or deletion of ornamentations, or the of the method is to analyze each input note determi- consolidation/fragmentation of notes. Apart from the ning (by means of the background musical knowled- tempo, other domain specific factors seem to play an ge) its role in the musical phrase it belongs to, identify important role in the way a melody is performed, such and retrieve (from the case-base of human perfor- as meter, and phrase structure. Tempo transformation mances) notes with similar roles, and finally, trans- is one of the audio post-processing tasks manually form the input note so that its expressive properties done in audio-labs. Automatizing this process may, (dynamics, rubato, vibrato, articulation, and attack) therefore, be of industrial interest. 6 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

7 Other applications of CBR to expressive music are generating an accompanying voice. In other words, those of Suzuki et al. (1999), and those of Tobudic you get an instrument that can be its own intelligent and Widmer (2003, 2004). Suzuki et al. (1999), use accompanist. Tod Machover, from MITs Media Lab, a082 examples cases of expressive performances to ge- developed such an hyper cello and the great cello pla- nerate multiple performances of a given piece with yer Yo-Yo Ma premiered, playing the hyper cello, a pie- Ramon Lpez de Mntaras Badia varying musical expression, however they deal only ce, composed by Tod Machover, called Begin Again with two expressive parameters. Tobudic and Widmer Again... at the Tanglewood Festival several years ago. (2003) apply instance-based learning (IBL) also to the problem of generating expressive performances. The Improvising music IBL approach is used to complement a note-level rule- Music improvisation is a very complex creative pro- based model with some predictive capability at the cess that has also been computationally modelled. It higher level of musical phrasing. More concretely, the is often referred to as composition on the fly and, IBL component recognizes performance patterns, of a therefore, it is, creatively speaking, more complex concert pianist, at the phrase level and learns how to than composition and it is probably the most com- apply them to new pieces by analogy. The approach plex of the three music activities surveyed here. An produced some interesting results but, as the authors early work on computer improvisation is the Flavours recognize, was not very convincing due to the limita- Band system of Fry (1984). Flavours Band is a proce- tion of using an attribute-value representation for the dural language, embedded in LISP, for specifying jazz phrases. Such simple representation cannot take into and popular music styles. Its procedural representa- account relevant structural information of the piece, tion allows the generation of scores in a pre-specified both at the sub-phrase level and at the inter-phrasal style by making changes to a score specification given level. In a subsequent paper, Tobudic and Widmer as input. It allows combining random functions and (2004), succeeded in partly overcoming this limita- musical constraints (chords, modes, etc.) to genera- tions by using a relational phrase representation. te improvisational variations. The most remarkable Widmer et al. (2009) describe a computer program result of Flavours Band was an interesting arrange- that learns to expressively perform classical piano ment of the bass line, and an improvised solo, of John music. The approach is data intensive and based on Coltranes composition Giant Steps. statistical learning. Performing music expressively cer- GenJam (1994) builds a model of a jazz musician tainly requires high levels of creativity, but the authors learning to improvise by means of a genetic algorithm. take a very pragmatic view to the question of whether A human listener plays the role of fitness function by their program can be said to be creative or not and rating the offspring improvisations. Papadopoulos claim that creativity is in the eye of the beholder. In and Wiggins (1998) also used a genetic algorithm to fact, the main goal of the authors is to investigate and improvise jazz melodies on a given chord progression. better understand music performance as a creative Contrarily to GenJam, the program includes a fitness human behaviour by means of AI methods. function that automatically evaluates the quality of The possibility for a computer to play expressively the offspring improvisations rating eight different as- is a fundamental component of the so-called hyper- pects of the improvised melody such as the melodic instruments. These are instruments designed to aug- contour, notes duration, intervallic distances between ment an instrument sound with such idiosyncratic notes, etc. nuances as to give it human expressiveness and a Franklin (2001) uses recurrent neural networks to rich, live sound. To make an hyper-instrument, take learn how to improvise jazz solos from transcriptions a traditional instrument, like for example a cello, and of solo improvisations by saxophonist Sonny Rollins. A connect it to a computer through electronic sensors in reinforcement learning algorithm is used to refine the the neck and in the bow, equip also with sensors the behaviour of the neural network. The reward function hand that holds the bow and program the computer rates the system solos in terms of jazz harmony crite- with a system similar to SaxEx that allows to analyse ria and according to Rollins style. the way the human interprets the piece, based on the score, on musical knowledge and on the readings of The lack of interactivity, with a human improviser, the sensors. The results of this analysis allow the hy- of the above approaches has been criticized (Thom, per-instrument to play an active role altering aspects 2001) on the grounds that they remove the musician such as timbre, tone, rhythm and phrasing as well as from the physical and spontaneous creation of a me- 7 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

8 lody. Although it is true that the most fundamental The discovery process of BACON is not a random characteristic of improvisation is the spontaneous, one and it could not be because the space of possible real-time creation of a melody, it is also true that in- functions to try is not finite and even if the search were a082 teractivity was not intended in these approaches and limited to a finite subset, any useful scientific domain nevertheless they could generate very interesting would be too large to allow random search. BACON Computational Creativity improvisations. Thom (2001) with her Band-out-of-a- uses several heuristics for searching selectively. First, Box (BoB) system addresses the problem of real-time starts with simple functions (as the linear function), interactive improvisation between BoB and a human then proceeds with more complex ones that are for- player. In other words, BoB is a music companion med by multiplying or dividing pairs of functions. Se- for real-time improvisation. Thoms approach follows cond, BACON uses data to guide the selection of the Johnson-Lairds (1991) psychological theory of jazz next function to try. More precisely, it notices if one improvisation. This theory opposes the view that im- variable increases or decreases monotonically with provising consists of rearranging and transforming respect to another. If it increases, it will test whether pre-memorized licks under the constraints of a har- the ratios of the values of the variables are constant. mony. Instead he proposes a stochastic model based If it decreases, it will test whether the products are on a greedy search over a constrained space of possi- constant. The main point is that BACON selects the ble notes to play at a given point in time. The very im- next function to test depending on how previously portant contribution Thom makes is that her system tried functions fit the data. Third, in situations in- learns these constraints, and therefore the stochastic volving more than two variables, BACON follows the model, from the human player by means of an unsu- well-known experimental procedure of changing one pervised probabilistic clustering algorithm. The lear- independent variable at a time. Having found condi- ned model is used to abstract solos into user-specific tional dependencies among small sets of variables, it playing modes. The parameters of that learned model explores the effects of altering other variables. are then incorporated into a stochastic process that With these simple means, and supplied with the generates the solos in response to four bar solos of actual data used by the original discoverers, BACON the human improviser. BoB has been very successfully rediscovered Keplers Third Law of planetary motion, evaluated by testing its real-time solo tradings in two Ohms Law of electric current and resistance, Blacks different styles, that of saxophonist Charlie Parker, Law of temperature equilibrium for mixtures of li- and that of violinist Stephane Grapelli. quids and many others. Another remarkable interactive improvisation sys- In validating BACON as a theory of human discovery, tem was developed by Dannenberg (1993). The diffe- Herbert Simon pointed out that BACON, interestingly rence with Thoms approach is that in Dannenbergs enough, initially arrived at the same erroneous square system, music generation is mainly driven by the law that Kepler himself had initially formulated, that composers goals rather than the performers goals. is the period of revolution of the planets varied as Wessels (1998) interactive improvisation system is the square of their distance to the Sun. However, closer to Thoms in that it also emphasizes the accom- BACON rejected it because it did not fit the data well paniment and enhancement of live improvisations. enough, and went on to discover the correct law. If BACON had used a larger error tolerance parameter it COMPUTATIONAL CREATIVITY IN SCIENCE would have also made Keplers mistake. According to BACON (Langley et al., 1987) is a representative Simon BACON shows that theories of inspiration are example of a program capable of rediscovering im- constructed and tested in exactly the same manner portant scientific laws using the generate and test as other scientific theories, that is, scientists need mechanism. BACON takes as inputs non-interpreted not to be seized by god to discover new laws. This numerical data and, when successful, produces scien- is true; Simon proceeds, even in discovering new con- tific laws that fit the data. Before proceeding with BA- cepts as BACON also shows. Indeed, using the heu- CON operational details, it is important to notice that ristic that when it discovers that there is an invariant the fact that it rediscovers known laws does not relation in the interaction between two or more ele- preclude its interest as a computational model of a ments in a situation, it should assign a new property creative process since, in principle, there is no reason to the elements, and measure its magnitude by the to believe that the cognitive processes involved in a relative strength of each elements action, BACON re- genuine discovery are different from those involved discovered the concept of inertial mass after noticing in a rediscovery. 8 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

9 that when pairs of bodies collide, the ratio of accele- involved, the number of reasons for suggesting the rations of any given pair is always the same. To do so, task and their worth and also the type of task. The top according to Simon, BACON defined a new property task is chosen relevant rules are gathered and applied a082 P and assigned a value 1 to that property for body resulting possibly in new concepts and new tasks. A and a value inversely proportional to the magnitude AM rediscovered natural numbers, addition, mul- Ramon Lpez de Mntaras Badia of their accelerations in collisions with A to the other tiplication, prime numbers, the prime factorization bodies. This procedure turns out to be a quite general theorem and the, yet unproven, Goldbachs conjectu- heuristic for discovering new concepts and has been re (any even number greater than two is the sum of used by BACON in discovering the concepts of specific two different primes) but soon reached its limits and heat, refractive index, voltage, molecular weight and failed to produce other interesting concepts. Instead atomic weight among others. Again, inspiration turns it proposed many boring tasks. According to Lenat, out to be a by-product of ordinary heuristic search in such early limitation was due to the fixed nature of Simons saying. the heuristics that worked well with simple concepts In my opinion, Simon was too optimistic. BACON such as sets but did not work with higher-level con- is too inflexible to be an acceptable computational cepts such as numbers. With the aim of overcoming model of scientific discovery. Indeed, its behaviour is this limitation, he extended AM to automatically mo- completely determined by its fixed set of heuristics dify its heuristic rules. The result was very disappoin- that guide an ad-hoc process of function approxima- ting since a large number of useless rules were crea- tion. There is no flexibility in the representation. To ted. According to Lenat (Rowe and Partridge, 1993), improve this, other programs such as GLAUBER (Lan- the reason was that many heuristics generated new gley et al., 1987) were written. These programs are concepts by syntactic manipulation of void concepts. able to induce structural and explanatory models of When these were definitions of mathematical con- certain phenomena. In particular GLAUBER examines cepts they were implemented as pieces of Lisp code qualitative data and produces classifications that in- and a minor modification to this Lisp code usually pro- duce general laws. For example, given a series of re- duced meaningful results. However, when similar mo- sults of chemistry experiments such as: difications were made to the Lisp code representing a heuristic rule, the results were meaningless. This (reacts input (C1H, NaOH) output (NaCl)) is because heuristics operate at a much higher level GLAUBER induces the law: than Lisp and many lines of Lisp code are required to implement each heuristic. Lenat, therefore, decided (reacts input (acid, alkali) output (salt)) that a new representation language at a higher level Another well known scientific discoverer is the Au- than Lisp was needed and he developed EURISKO tomated Mathematician or AM (Lenat, 1983), a pro- (Lenat, 1983) in which heuristics were represented gram that (re)discovered mathematical concepts in as frames with slots containing small pieces of Lisp number theory based on a hierarchy of about 100 code. The idea was to produce meaningful changes in basic concepts (sets, ordered pairs, basic operations the heuristics by mutations in the values of the slots. such as union, intersection, etc.) and using some 250 EURISKO was applied to VLSI design, space-ship fleet heuristic rules to guide the discovery process. Each design and number theory and it worked quite well concept is represented by a set of slots containing when interacting with a user that could eliminate ob- information such as definition, examples, specializa- viously bad heuristics. However it did not go further tions, worth, and in the case of operations its domain than AM in number theory. It seems that, once more, and range. The heuristic rules are of four different ty- the heuristics were not good enough. Going further pes: Fill, check, suggest and interest. Thus AM can use in number theory requires dealing with concepts in interest rules to evaluate the interest of the concepts many other mathematical domains such as algebra, it discovers and, therefore, guide the search towards geometry, graph theory, etc. and neither AM nor EU- more promising concepts. Fill rules try to fill the exam- RISKO could deal with such concepts. ples slot with examples of a concept. Check rules verify Besides, AM has been criticized by Ritchie and Han- correctness of the examples and also look for regulari- na (1984) on other grounds. According to Ritchie and ties. Suggest rules are considered when the program Hanna, many rules contain special purpose hacks, is running low on interesting things to do. Overall, AM never described by Lenat, that seem to have been is controlled by an agenda which maintains a list of written for the specific purpose of making the most tasks sorted according to the interest of the concepts 9 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

10 interesting discoveries. The precise extent of AM One very creative answer of COPYCAT was wyz. This creativity is, therefore, rather unclear. answer involves simultaneously relating left with right, first letter of the alphabet with last letter of a082 The major limitation of all these programs is that the alphabet, and successor with predecessor. In what to look for is built into the heuristics and that spite of operating in a very limited domain, COPYCAT such heuristics are too inflexible. A creative idea also Computational Creativity shows interesting features of creative processing. involves breaking rules or dropping constraints, and a mechanism to do so is that of reasoning by analogy. Other programs such as MECHEM (Valds-Prez, An excellent example of a human discovery, based on 1995) incorporate sophisticated techniques of reaso- analogy and constraints dropping, is that of Kekules ning by analogy and constraint satisfaction that allow discovery of the benzene-ring structure (other well them to reason about the structural transformations known examples of constraint dropping are non- that take place in chemically reacting molecules with Euclidean geometry and Schoenbergs non tonal mu- the aim of eliciting the internal, non observable, me- sic). Kekule described his discovery as follows (Boden, chanisms involved in the reactions based on empirical 1991): evidence. MECHEM has discovered and explained me- I turned the chair to the fire and dozed. Again the chanisms in the hydrogenolosis of methane that coin- atoms were gamboling before my eyes... (My mental cide fully with results published just a couple of years eye) could distinguish larger structures, of manifold earlier in the Journal Catalysis Today. Understanding conformation; long rows, sometimes more closely the internal mechanisms involved in chemical reac- fitted together; all twining and twisting in snakelike tions has an enormous practical interest since that motion. But look! What was that? One of the snakes knowledge can suggest better ways of controlling the had seized hold of its own tail, and the form whirled reactions. It has been conjectured that programs like mockingly before my eyes. As if by a flash of lightning MECHEM may well be necessary to understand the I awoke. mechanisms implicit in complex chemical reactions. This vision was the origin of his discovery that the Bob Holmes (1996) reported the development of benzene molecule was a ring and not a chain. The a Creativity Machine developed by the materials analogy between snakes and molecules was certainly scientist Steve Thaler. Among other applications, it a fundamental aspect of the discovery that triggered discovered ultra-hard materials. This system is based the dropping of the constraint that molecules should on a large network with inputs and outputs repre- be chains (open curves) and allowing them to be clo- senting every possible quantum state for every elec- sed curves (rings). tron in every atom of a molecule. Thaler trained this Psychologists have studied analogy for a long time network by showing it about 200 examples of two- and one of the main conclusions was that analogy is element molecules, such as water and iron oxide, context-sensitive. Indeed, analogy in the context of to teach it plausible combinations and proportions. poetry is not the same as in the context of science. Holmes reports that the machine correctly identified COPYCAT (Mitchell and Hofstadter, 1990) was the first known ultra-hard materials such as boron nitride and computational model of analogy and is heavily based boron carbide, even though it had never seen these on human psychology and particularly on context sen- during training. It also proposed a material, C3N4 sitivity. This system generates many candidate analo- that a group of theoreticians from Harvard had al- gies but only those that are contextually appropria- most simultaneously suggested as a likely ultra-hard te are kept. COPYCAT constructs analogies between material. Holmes adds that the system also pointed strings of letters. Let us see an example taken from out several untested polymers of boron, beryllium, Boden (1994). COPYCAT is given as input that the or carbon doped with small amounts of hydrogen. string abc rewrites into abd, and then is asked what The Creativity Machine was licensed to a company the string mrrjjj will rewrite into. There are several to develop new ultra-hard materials and high-tem- acceptable answers to this depending on the context perature superconductors. that COPYCAT has considered. For instance, if the con- Another recent very remarkable achievement is a text favours successors over repetitions, then one robot-scientist, called ADAM (King et al. 2004), which answer would be mrrjjk. But if it favours repetitions conducted experiments on yeast using AI techniques. then the answer is mrrjjjj. Another interesting analogy The goal of these experiments was to determine the (Boden, 1994) found by COPYCAT is the following one: function of several gene knockouts by varying the If abc rewrites into abd, what will xyz rewrite into? quantities of nutrient provided to the yeast. The robot 10 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

11 used a machine learning technique known as induc- agents. The lack of intentionality is a direct conse- tive logic programming to select those experiments quence of Searles Chinese room argument (Searle, that could discriminate between different hypothe- 1980), which states that computer programs can only a082 ses. Feedback on each experiment was provided by perform syntactic manipulation of symbols but are data reporting yeast survival or death. The most ac- unable to give them any semantics. This criticism is Ramon Lpez de Mntaras Badia curate robot strategy outperformed humans doing based on an erroneous concept of what a computer the same task. Stephen Muggleton, one of the desig- program is. Indeed, a computer program does not ners of ADAM, predicts the development of the first only manipulate symbols but also triggers a chain of micro-fluidic robot scientist within the next years. A cause-effect relations inside the computer hardware micro-fluidic robot scientist, according to Muggleton and this fact is relevant for intentionality since it is ge- (2006), will combine active learning and autonomous nerally admitted that intentionality can be explained experimentation with micro-fluidic technology (Flet- in terms of causal relations. However, it is also true cher et al., 2001). He argues that nowadays scientists that existing computer programs lack too many rele- can already build miniaturized laboratories on a chip vant causal connections to exhibit intentionality, but using micro-fluidics and since these chips can perform perhaps future, possibly anthropomorphic, embo- chemical synthesis and testing at high speed, one can died artificial intelligences --that is agents equipped imagine miniaturizing the robot-scientist technology not only with sophisticated software but also with with the goal of reducing the experimental cycle time different types of advanced sensors allowing to inte- from hours to milliseconds. Furthermore, Muggleton ract with the environment-- may have sufficient causal speculates that more flexibility could be added to the connections to exhibit intentionality. micro-fluidic machines by developing what he calls a Regarding social rejection, the reasons why we are Chemical Turing Machine. The Chemical Turing Ma- so reluctant to accept that non biological agents can chine, according to Muggleton, would be a univer- be creative (even biological ones as it is the case with sal processor capable of performing a broad range of Nonja, a 20 years old painter from Vienna whose chemical operations on both the reagents available to abstract paintings had been exhibited and apprecia- it at the start and those chemicals it later generates. ted in art galleries but that after knowing that she The machine would automatically prepare and test was an orang-utan from the Vienna Zoo, her work chemical compounds but it would also be program- was much less appreciated!) is that they do not have mable, thus allowing much the same flexibility as a a natural place in our society of human beings and a real chemist has in the lab. decision to accept them would have important social Similarly to a standard Turing Machine, an automa- implications. It is therefore much simpler to say that ton, introduced by Alan Turing to give a mathematica- they appear to be intelligent, creative, etc. instead of lly precise definition of algorithm (Turing, 1936, 1938), saying that they are. In a word, it is a moral but not a which consisted of an infinite tape and a set of very scientific issue. A third reason for denying creativity simple rules for moving the tape and manipulating to computer programs is that they are not conscious the symbols that contains, a Chemical Turing Machine of their accomplishments. However I agree with many would be an automaton connected to a conveyor belt AI scientists in thinking that the lack of consciousness containing a series of flasks: the chemical Turing ma- is not a fundamental reason to deny the potential for chine can move the conveyor to obtain distant flasks, creativity or even the potential for intelligence. After and can mix and make tests on local flasks. all, computers would not be the first example of un- conscious creators, evolution is the first example as CONCLUDING REMARKS: APPARENTLY OR REALLY Stephen Jay Gould (1996) brilliantly points out: If CREATIVE? creation demands a visionary creator, then how does blind evolution manage to build such splendid new Margaret Boden pointed out that even if an arti- things as ourselves? ficially intelligent computer would be as creative as Bach or Einstein, for many it would be just apparently creative but not really creative. I fully agree with her ACKNOWLEDGEMENTS in the two main reasons for such rejection. These rea- This research has been partially supported by the sons are: the lack of intentionality and our reluctance EU FP7 PRAISE project #318770 and by the 2009-SGR- to give a place in our society to artificially intelligent 1434 Grant from the Generalitat de Catalunya. 11 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

12 references Arcos, J. L.; Lpez de Mntaras, R. (2001). Cope, D. (1987). Experiments in Music In- Friberg, A.; Sunberg, J.; Fryden, L. (2000). a082 An Interactive Case-Based Reasoning telligence. In Proceedings of the 1987 Music From Motion: Sound Level En- Approach for Generating Expressive International Computer Music Confe- velopes of Tones Expressing Human Music. Applied Intelligence, 14 (1), pp. rence. San Francisco, Calif.: Internatio- Locomotion. Journal on New Music Re- Computational Creativity 115-129. nal Computer Music Association. search, 29 (3), pp. 199-210. Arcos, J. L.; Lpez de Mntaras, R.; Serra, X. Cope, D. (1990). Pattern Matching as an Fry, C. (1984). Flavors Band: A language for (1998). Saxex: A Case-Based Reasoning engine for the computer simulation of specifying musical style. Reprinted in System for Generating Expressive Musi- musical style. In Proceedings of the S.M. Schwanauer and D.A. Levitt (eds.) cal Performances. Journal of New Music 1990 International Computer Music (1993), Machine Models of Music (pp. Research, 27 (3), pp. 194-210. Conference. San Francisco, Calif.: Inter- 427-451). Cambridge, Mass.: The MIT Bentley, P. J.; Corne, D.W. (eds.) (2002). national Computer Music Association. Press. Creative Evolutionary Systems. Morgan Dannenberg, R. B. (1993). Software design Grachten, M.; Arcos, J. L.; Lopez de Man- Kaufmann. for interactive multimedia performan- taras, R. (2004). TempoExpress, a CBR Bharucha, J. (1993). MUSACT: A connectio- ce. Interface, 22 (3), pp. 213-218. Approach to Musical Tempo Transfor- nist model of musical harmony. In S. M. mations. In P. Funk and P. A. Gonzalez Dannenberg, R. B.; Derenyi, I. (1998). Schwanauer and D.A. Levitt (eds.), Ma- Calero, (eds.), Proceedings of the 7th Combining instrument and performan- chine Models of Music (pp. 497-509). European Conference on Case-Based ce models for high quality music synthe- Cambridge, Mass.: The MIT Press. Reasoning), Lecture Notes in Artificial sis. Journal of New Music Research, 27 Intelligence, vol. 3155, pp. 601-615. Biles, J. A. (1994). GenJam: A genetic algo- (3), pp. 211-238. rithm for generating Jazz solos. In Pro- Gould, S. J. (1996). Creating the Creators. Dartnall, T. (ed.) (1994). Artificial Intelligen- ceedings of the 1994 International Com- Discover Magazine, October, pp. 42-54. ce and Creativity. Kluwer Academic Pub. puter Music Conference. San Francisco, Hiller, L.; Isaacson, L. (1958). Musical com- Desain, P.; Honing, H. (1993). Tempo cur- Calif., International Computer Music position with a high-speed digital com- ves considered harmful. In J. D. Kramer Association. puter (1958). Reprinted in S.M. Schwa- (ed.), Time in contemporary musical Boden, M. (1991). The Creative Mind: nauer and D.A. Levitt (eds.) (1993), thought. Contemporary Music Review, Myths and Mechanisms. New York: Ba- Machine Models of Music (pp. 9-21). 7 (2). sic Books. Cambridge, Mass.: The MIT Press. Ebcioglu, K. (1993). An expert system for Boden, M. (ed.) (1994). Dimensions of Crea- Holmes, B. (1996). Steve Thalers Creativi- harmonizing four-part chorales. In S.M. tivity. Cambridge, Mass.: The MIT Press. ty Machine. New Scientist, Vol. 2013, Schwanauer and D.A. Levitt (eds.), Ma- pp. 22-24. Boden, M. (2009). Computers Models of chine Models of Music (pp. 385-401). Creativity. AI Magazine, 30 (3), pp. 23- Cambridge, Mass.: The MIT Press. Hrnel, D.; Degenhardt, P. (1997). A neu- 34. ral organist improvising Baroque-style Feulner, J. (1993). Neural Networks that melodic variations. In Proceedings of Bresin, R. (1998). Artificial neural networks learn and reproduce various styles of the 1997 International Computer Music based models for automatic performan- harmonization. In Proceedings of the Conference (pp. 430-433). San Francis- ce of musical scores. Journal of New 1993 International Computer Music co: International Computer Music Asso- Music Research, 27 (3), pp. 239-270. Conference. San Francisco: Internatio- ciation. nal Computer Music Association. Bresin, R. (2001). Articulation rules for Hrnel, D.; Menzel, W. (1998). Learning automatic music performance. In Pro- Fletcher, P.; Haswell, S.; Watts, P. & Zhang, musical strucure and style with Neural ceedings of the 2001 International Com- X. (2001). Lab-on-a-Chip Micro Reac- Networks. Journal on New Music Re- puter Music Conference. San Francisco, tors for Chemical Synthesis. Dekker search, 22 (4), pp. 44-62. Calif.: International Computer Music Encyclopedia of Nanoscience and Nano- Association. technology. Johnson, M. L. (1992). An expert system for the articulation of Bach fugue me- Buchanan, B. G. (2001). Creativity at the Franklin, J. A. (2001). Multi-phase learning lodies. In D.L. Baggi (ed.), Readings in Metalevel: AAAI-2000 Presidential for jazz improvisation and interaction. Computer Generated Music (pp. 41-51), Address, AI Magazine, 22 (3), pp. 13-28. In Proceedings of the Eight Biennial Los Alamitos, Calif.: IEEE Press. Symposium on Art and Technology. Canazza, S.; De Poli, G.; Roda, A.; Vidolin, Johnson-Laird, P. N. (1991). Jazz improvi- A. (1997). Analysis and synthesis of Friberg, A. (1995). A quantitative rule sys- sation: A theory at the computational expressive intention in a clarinet per- tem for musical performance. (unpublis- level. In P. Howell, R. West, and I. Cross, formance. In Proceedings of the 1997 hed PhD dissertation). KTH, Stockholm. (ed.), Representing Musical Structure. International Computer Music Con- Friberg, A.; Bresin, R.; Fryden, L.; Sunberg, London, UK and San Diego, Calif.: Aca- ference (pp. 113-120). San Francisco, J. (1998). Musical punctuation on the demic Press. Calif.: International Computer Music microlevel: automatic identification and Association. Kendall, R. A., Carterette, E. C. (1990). The performance of small melodic units. communication of musical expression. Colton, S.; Lpez de Mntaras, R.; Stock, Journal of New Music Research, 27 (3), Music Perception, 8 (2), pp. 129. O. (2009). Computational Creativity: pp. 271-292. Coming of Age. Special Issue of AI Ma- gazine, vol. 30 no. 3 pp. 11-88. 12 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

13 King, R. D.; Whelan, K. E.; Jones, F. M.; matic harmonization. In ECAI98 Wor- ration of Musical Expression. In Pro- Reiser, P. G. K.; Bryant, C. H.; Mu- kshop on Constraint Techniques for Ar- ceedings of the 16th International Joint ggleton, S. H.; Kell, D. B.; Oliver, S. G. tistic Applications. Brighton, UK. Conference on Artificial Intelligence, (2004). Functional genomic hypothe- Morgan Kaufmann pp. 642-648. Papadopoulos, G.; Wiggins, G. (1998). A a082 sis generation and experimentation genetic algorithm for the generation Thom, B. (2001). BoB: An improvisatio- by a robot scientist. Nature, Vol. 427, of Jazz melodies. In Proceedings of the nal music companion. (unpublished Ramon Lpez de Mntaras Badia pp. 247-252. SteP98 Conference, Finland. PhD dissertation). School of Computer Langley, P.; Simon, H. A.; Bradshaw, G. L.; Science, Carnegie-Mellon University. Partridge, D.; Rowe, J. (1994). Computers Zytkow, J. M. (1987). Scientific Disco- and Creativity. Intellect Books. Tobudic, A.; Widmer, G. (2003). Playing very, Computational Explorations of the Mozart Phrase by Phrase. In K.D. As- Creative Mind. Cambridge, Mass.: The Rader, G. M. (1974). A method for compo- hley and D.G. Bridge (eds.), Proceedings MIT Press. sing simple traditional music by compu- of the 5th International Conference on ter. Reprinted in S.M. Schwanauer and Lenat, D. B. (1983). The role of heuristics Case-Based Reasoning. Lecture Notes D.A. Levitt (eds.) (1993), Machine Mo- in learning by discovery: Three case stu- in Artificial Intelligence, vol. 3155, pp. dels of Music (pp. 243-260). Cambridge, dies. In Machine Learning: An Artificial 552-566. Mass.: The MIT Press. Intelligence Approach. Tioga. Tobudic, A.; Widmer, G. (2004). Case-Ba- Ritchie, G. D.; Hanna, F. K. (1984). AM: A Lerdahl, F.; Jackendoff, R. (1983). An over- sed Relational Learning of Expressive case study in AI methodology. Artificial view of hierarchical structure in music. Phrasing in Classical Music. In P. Funk Intelligence, vol. 23, pp. 249-268. Music Perception, vol. 1, pp. 229-252. and P. A. Gonzalez Calero (eds.), Procee- Rothgeb, J. (1969). Simulating musical dings of the 7th European Conference Levitt, D. A. (1993). A Representation for skills by digital computer. Reprinted on Case-Based Reasoning. Lecture No- Musical Dialects. In S.M. Schwanauer in S.M. Schwanauer and D.A. Levitt tes in Artificial Intelligence, vol. 3155, and D.A. Levitt (eds.), Machine Models (eds.) (1993), Machine Models of Mu- pp. 419-433. of Music (pp. 455-469). Cambridge, sic (pp. 157-164). Cambridge, Mass.: Mass.: The MIT Press. Turing, A. M. (1936). On Computable The MIT Press. Numbers, with an Application to the Lopez de Mantaras, R.; Arcos, J. L. (2002). Rowe, J.; Partridge, D. (1993). Creativity: A Entscheidungsproblem. Proceedings of AI and Music: From Composition to Ex- survey of AI approaches. Artificial Inte- the London Mathematical Society, vol. pressive Performance. AI Magazine, 23 lligence Review, vol. 7, pp. 43-70. 42, pp. 230265. (3), pp. 43-57. Sabater, J.; Arcos, J. L.; Lopez de Mantaras, Turing, A. M. (1938). On Computable Minsky, M. (1981). Music, Mind, and Me- R. (1998). Using Rules to Support Case- Numbers, with an Application to the aning. Reprinted in S.M. Schwanauer Based Reasoning for Harmonizing Me- Entscheidungsproblem: A correction. and D.A. Levitt (eds.) (1993), Machine lodies. In AAAI Spring Symposium on Proceedings of the London Mathemati- Models of Music (pp. 327-354). Cam- Multimodal Reasoning, 147-151. Menlo cal Society vol. 43, pp. 544546. bridge, Mass.: The MIT Press. Park, Calif.: American Association for Ar- Turing, A. M. (1950). Computing Machi- Mitchell, M.; Hofstadter, D. (1990). The tificial Intelligence. nery and Intelligence. Mind, vol. LIX no. Emergence of Understanding in a Schwanauer, S. M. (1993). A learning ma- 236, pp. 433-460. Computer Model of Concepts and chine for tonal composition. In S.M. Analogy Making. Physica D, vol. 42, Valds-Prez, R. E. (1995). Machine dis- Schwanauer and D.A. Levitt (eds.), Ma- pp. 322-334. covery in chemistry: New results. Ar- chine Models of Music (pp. 511-532). tificial Intelligence Journal, vol. 74, pp. Moorer, J. A. (1972). Music and computer Cambridge, Mass.: The MIT Press. 191-201. composition. Reprinted in S.M. Schwa- Schwanauer, S. M., Levitt, D. A. (eds.) nauer and D.A. Levitt (eds.) (1993), Ma- Wessel, D.; Wright, M.; Kahn, S. A. (1998). (1993). Machine Models of Music. Cam- chine Models of Music (pp. 167-186). Preparation for improvised performan- bridge, Mass.: The MIT Press. Cambridge, Mass.: The MIT Press. ce in collaboration with a Khyal singer. Searle, J. (1980). Minds, Brains and Pro- In Proceedings of the 1998 International Morales-Manzanares, R.; Morales, E. F.; grams. Behavioral and Brain Sciences, Computer Music Conference. San Fran- Dannenberg, R.; Berger, J. (2001). SI- 3 (3) pp. 417-457. cisco, Calif.: International Computer CIB: An Interactive Music Composition Simon, H. A.; Sumner, R. K. (1968). Patterns Music Association. System Using Body Movements. Com- puter Music Journal, 25 (2), pp. 25-36. in Music. Reprinted in S.M. Schwan- Widmer, G.; Flossmann, S.; Grachten, M. auer and D.A. Levitt (eds.) (1993), Ma- (2009). YQX Plays Chopin. AI Magazi- Muggleton, S. H. (2006). Exceeding human chine Models of Music (pp. 83-110). ne, 30 (3), pp. 35-48. limits. Nature, Vol. 440, pp. 409-410. Cambridge, Mass.: The MIT Press. Woods, W. (1970). Transition network gra- Pachet, F.; Roy, P. (1998). Formulating Suzuki, T.; Tokunaga, T.; Tanaka, H. (1999). mmars for natural language analysis. constraint satisfaction problems on A Case-Based Approach to the Gene- Communications of the ACM, 13 (10), part-whole relations: The case of auto- pp. 591-606. 13 ARBOR Vol. 189-764, noviembre-diciembre 2013, a082. ISSN-L: 0210-1963 doi: http://dx.doi.org/10.3989/arbor.2013.764n6005

Load More