"A Scheme for Learning and Behaviour",
C.M. Witkowski
1994 AISB Workshop "Models
or Behaviours - Which Way Forward for Robotics?"
University of Leeds,
11-13 April 1994, 5pp.
ABSTRACT
Learning ability is widely recognised as fundamental to our notion of
intelligence, whether natural or artificial. The overall goal of the SLAB
(Scheme for Learning and Behaviour) research is to investigate the degree to
which robot behaviour may be acquired without the intervention of a supervisor
cognisant of the "correct" behaviour, and without any direct
indication of the correctness (the learning "anchor") pre-coded into
the system. Depending on the strictness with which these restrictions are
applied the research problem of generating novel and purposeful robot behaviour
has ranged from difficult to extremely intractable. SLAB primarily approaches
these problems by the creation, validation and modification of a number of
different classes of prediction.SLAB investigates the internal mechanisms and
structures required for an individual system's performance. It is not, at
present, targeted at bulk population behaviours or performance changes over a
population, as with genetic algorithms and A-Life investigations.
Directly
executed "behaviours" have long been recognised to be capable of
generating complex, interesting and effective activity. In many instances it can
prove difficult to determine by observation alone whether behaviour is generated
directly, or derived from a deeper model. In the industrial application of
robotics the exclusive use of direct behaviour programming often proves highly
desirable in terms of predictability, maintainability, and from a safety
standpoint. Directly executed behaviour finds an important place in SLAB - in
providing seed behaviours and in making learnt behaviours efficient - but
ultimately direct behaviours alone place unacceptable restrictions on the
behavioural complexity and learning potential of the overall system.
In
SLAB the robot and its external environment are modelled to the extent that
sensor conditions and events are divided into those which may be predicted, and
those which the robot may actively define as a goal and as a consequence
actively pursue. In implementation, as the rule based programming system
MONOLITH, a surprisingly complex system architecture was found to be required.
While sharing a common syntax and structure, rules within MONOLITH are
specialised to perform a dozen different activities all related to internal and
external behaviour. Sensor input is also essential to the learning and behaving
activities. Without continued sensor input robot behaviour is restricted and
learning is reduced to an essentially limited process of maturation.