Information Technology Reference
In-Depth Information
inexperienced improvisors; more experienced improvisors exhibit some degree of
coherence in their approach. Regularity, as captured by consistency and the ability
to plan ahead, is measurable by the relationship of the current output to previous
and future outputs.
The musical elements in question have not been defined, and a prescription of
how to make the measurements has not been detailed. The process is dependent on
the level of description, and it may be that several levels are needed. At the level of
note events, for example, the elements are phrases, and comparisons can be made
using a similarity measure based on a definition of the distance between two phrases
(how closely the phrases shapes match each other, or the number of changes needed
to bring the phrases into agreement, for example). Comparisons between streams are
also possible using information-theoretic techniques such as mutual entropy; such
methods might be important where an appropriate level cannot be defined.
6.3.5 Artificial Intelligence
Artificial Intelligence (AI) offers various schemes that may prove fertile for Live Al-
gorithm research and strategies for developing functions
f
, as represented in general
in the wiring diagrams above.
Reasoning
can be based on a programmed rule set, or derived through training.
In the former, the Live Algorithm designer has access to the vast experience of sym-
bolic AI research that includes knowledge representation, problem solving, planning
and expert systems. Within this AI framework, the problem domain is represented
symbolically. The representation contains knowledge about the problem domain,
and the symbols are syntactically manipulated until a solution is found. The main
focus of the symbolic framework is on a suitable formal representation of the prob-
lem domain, the inclusion of domain specific knowledge and efficient algorithms.
Machine learning
is another major AI framework. The learning algorithm can be
based, for example, on a neural architecture or on Bayesian structures (e.g. Hidden
Markov Modelling). Responses are learnt over a sequence of test cases. The focus
is on the learning algorithm, the training set and the network architecture.
We can suppose that a human improviser can potentially refer to his/her own the-
oretic knowledge as well as her/his experiential knowledge of music making and of
group improvisation in particular. It would be inhibitive to deny similar advantages
to a Live Algorithm. Domain knowledge can be hard-wired into the Live Algorithm
and trial performances offer ideal test cases for learning algorithms. Since the defi-
nition of the Live Algorithm only makes reference to inferred behaviour and not to
any supposed mental states, the debate as to whether cognition is symbol manipula-
tion (computationalism) or dependent on a neural architecture (connectionism), or
indeed some other paradigm, is not relevant; rather, any technique can be requisi-
tioned in order to further the overall goals of Live Algorithm research.
As an alternative approach to reasoning or learning, we mention here the
dynam-
ical systems
framework which has already proven to be a rich source of ideas in
Live Algorithm research.
