%A K.A. Lantz
%T Uniform interfaces for distributed systems
%R Ph.D. thesis
%I University of Rochester, NY
%D May 1980

%A Lokendra Shastri
%T Evidential reasoning in semantic networks:
A formal theory and its parallel implementation
%R Ph.D. thesis
%I University of Rochester, NY
%D July 1985
%X ABSTRACT:
The problem of representing and utilizing a large body of knowledge is
fundamental to artificial intelligence. This thesis focuses on two
important issues related to this problem.
	1. An agent cannot maintain complete information about any
	   but the most trivial environment, and therefore, he must
	   be capable of reasoning with incomplete and uncertain
	   information.
	2. An agent must act in real-time. Human agents take a few
	   hundred milliseconds to perform a broad range of tasks,
	   and an agent endowed with artificial intelligence should
	   perform similar tasks in comparable time.
It is argued that the best way to cope with partial and incomplete
information is to adopt an evidential form of reasoning, wherein,
inference does not involve establishing the truth of a proposition but
instead, it involves finding the most likely hypothesis from among a
set of alternatives.
It is also argued that in order to satisfy the real-time constraint,
we must identify the kinds of inference that we must perform very
fast, and provide a c_o_m_p_u_t_a_t_i_o_n_a_l a_c_c_o_u_n_t of how
this limited class of inference may be performed in an acceptable time
frame.  This latter requirement prompts us to consider massively
parallel (connectionist) models of computation, in particular models
that do not require an interpreter.
Inheritance and categorization within a conceptual hierarchy are
identified as two operations tha humans perform very fast. It is
suggested that these operations are important because they seem to lie
at the core of intelligence and are precursors to more complex
reasoning.
The above concerns and proposed solutions lead to an evidential
framework for representing conceptual knowledge wherein the principle
of maximum entropy is applied to deal with uncertainty and
incompleteness.  It is demonstrated that the proposed framework offers
a uniform treatment of inheritance and categorization, and solves an
interesting class of inheritance and categorization problems,
including those that involve exceptions, multiple hierarchies, and
conflicting information. The proposed framework can be encoded as an
interpreter-free, massively parallel (connectionist) network, that can
solve inheritance and categorization problems in time proportional to
the depth of the conceptual hierarchy.
(Advisor: Jerome A. Feldman)