Autonomous Learning from the
Environment
I believe that intelligent behavior of any creature is ultimately
rooted in its physical abilities to perceive and act in its
environment. ``Our
idea of anything is our idea of its sensible effects.'' We say ice is
cold,
because we touch it and sense the effect. The secret of autonomous
learning,
I think, lies inside the process of how such sensible effects are
learned
by each individual system from its environment. Driven by this basic
belief,
I have identified three essential tasks for autonomous learning: Model
Abstraction, Model Application, and Integration, and
developed
a unified computational framework called Predict-Surprise-Revise.
This framework, when applied to concept learning tasks, yields
- A highly efficient incremental algorithm called CDL2 for learning
decision lists.
When applied to environment learning tasks, the framework produces
- An algorithm called LDE
for learning finite state machines, and
- A system called LIVE
that learns action-prediction rules from various simulated environments
(including puzzles, robot hands and eyes, psychological experiments on
child development, and historical experiments of scientific
discoveries).
A good summary of this line of research is included in my recent book Autonomous Learning from
the Enivronment, published by Computer Science Press.
The research also reveals several important open problems. It is
not clear at all how humans construct necessary mental features when
learning in a novel environment. Previous cognitive research assumes it
is done by the subconscious.
The LIVE system bypasses this problem by limiting the possible mental
constructors
to the minimal. Clearly, plenty of research needs to be done here. As a
related
problem, it is also not clear how high level cognitive models deal with
uncertainties
and errors of physical systems. I have been investigating the use of
Bayesian
probability theory as well as parallel computational models such as
neural
networks. I plan to conduct experiments on applying the above framework
and
its extensions to physical robots.
REFERENCES
Books
Shen, W-M.,
Autonomous Learning from
the Environment (355 pages), W. H. Freeman, Computer
Science Press, 1994. (Foreword by Herbert A. Simon)
Shen, W.M., Edited,
Learning Actions Models, AAAI Press, 1999.
Journal
Publications
- Shen,
WM. The Process of Discovery. Foundations of Science,
1(2), 1995.
- Shen,
WM. Discovery as
autonomous learning from the environment. Machine
Learning , 12, 143-156, 1993.
- Shen, WM.
and H.A.
Simon. Fitness requirements for scientific theories containing
recursive
theoretical terms. British Journal of Philosopy of Science, 44,
641-652,
1993.
- Shen, WM.
Discovering regularities from knowledge bases. International
Journal of Intelligent Systems, 7(7), 623--636, 1992.
- Shen, WM.
LIVE: An architecture for autonomous learning from the environment, ACM
SIGART Bulletin, Special issue on Integrated Cognitive
Architectures, 2(4),
151-155, 1992.
- Shen,
WM. Functional Transformation in AI Discovery Systems. Artificial
Intelligence , 41(3), 257-272, 1989.
Book
Chapters
- Shen, WM.
1994. Learning deterministic finite automata using local distinguishing
experiments. In Computational Learning Theory and Natural Learning
Systems, edited by T. Petsche and S. Judd and S. Hanson. MIT Press.
Conference Papers
- Shen, W.-M., Learning Finite Automata
Using Local Distinguishing Experiments, IJCAI 1993.
- Shen, W.-M., Complementary
Discrimination Learning with Decision Lists, AAAI 1992.