Information Technology Reference
In-Depth Information
previously known spatio-temporal objects in non-trivial ways to create novel pat-
terns.
According to the Formal Theory of Creativity , in the examples above, people at-
tempt to maximise essentially the same type of objective function or reward function
at various stages of their lives. Part of the reward is standard external reward as used
in many applications of Reinforcement Learning (RL) (Kaelbling et al. 1996 ), such
as positive reward for eating when hungry, or negative reward (pain) for bumping
into an obstacle. In addition to that, however, there is the intrinsic reward, or aes-
thetic reward, or pure fun, which a creative, subjective observer may extract from
some self-generated sequence of actions and observations by learning to encode it
more efficiently: the fun is proportional to the difference between how many com-
putational resources (storage space and time) he needs to encode the data sequence
before and after learning. A separate RL algorithm maximises expected fun by find-
ing or creating non-random, non-arbitrary data that soon becomes more predictable
or compressible in some initially unknown but learnable way, such as novel jokes,
songs, dances, paintings, or scientific observations obeying novel, unpublished laws.
In Sect. 12.3 we will formalise the basic principle. In Sect. 12.4 we discuss
our previous approximative implementations thereof: concrete examples of artificial
creative scientists or artists that learn to create action sequences yielding intrinsic
aesthetic rewards independent of human supervision. In Sect. 12.5 we summarise
why aesthetic reward can be viewed as the first derivative of subjective beauty in
the sense of elegance or simplicity. In Sect. 12.6 we describe the creation of a work
of Low-Complexity Art (Schmidhuber 1997c ) computable by a very short program
discovered through a search process modelled by the Formal Theory of Creativ-
ity. Next, however, we will first discuss relationships to previous ideas on curiosity,
creativity, and aesthetic reward.
12.2 Relation to Previous, Less Formal Work
Much of the work on computational creativity described in this topic uses reward
optimisers that maximise external reward given by humans in response to artistic
creations of some improving computational pattern generator. This chapter, how-
ever, focuses on unsupervised creative and curious systems motivated to make novel,
aesthetically pleasing patterns generating intrinsic reward in proportion to learning
progress.
Let us briefly discuss relations to previous ideas in this vein. Two millennia ago,
Cicero already called curiosity a “passion for learning”. Section 12.3 will formalise
this passion such that one can implement it on computers, by mathematically defin-
ing reward for the active creation of patterns that allow for compression progress or
prediction improvements.
In the 1950s, psychologists revisited the idea of curiosity as the motivation for
exploratory behaviour (Berlyne 1950 ; 1960 ), emphasising the importance of nov-
elty (Berlyne 1950 ) and non-homeostatic drives (Harlow et al. 1950 ). Piaget ( 1955 )
Search WWH ::




Custom Search