Copyright © 1997, All Rights Reserved.
All materials contained hereafter are copyright property of John C. Wang, no parts can be reproduced without prior consent from the author.
John C. Wang is a graduate student at the National Taiwan University, Taiwan. He is a member of the Intelligent Robotics Lab at NTU, author of the UniBase Network Programming Framework, founder of the Ideae Group, and an active member of the intelligent agent research community. He can be reached at http://ideae.csie.ntu.edu.tw/jcwang/.
Edmund Husserl's phenomenology in twentieth century philosophy takes experience as the primary grounds for establishing the universal truth. Here we refer to a phenomenological agent as one that can be described solely in terms of external observations made on the agent.
Human beings often develop a model for an external agent through obseravtions and interactions, recording the experiences in the process. Humans are more comfortable working with predictable, describable, and thus phenomenological agents. The world we live in, the family members we are brought up with, and the value of the currency we trade in, are all phenomenological in this respect. For this reason we are comfortable living in this world the way it is.
Acts that are considered intelligent are often those of the predictable person, and are often desirable for an interactive intelligent agent to have. Examples include the ability of natural language processing, acts of intuition and instinct, and the capacity for friendliness, emotions, etc.
Current practice in artificial intelligence may approach intelligence from cognitive science, using artificial models of human intelligence constructed from empirical evidence. Another camp of researchers drops the human issue for the moment and takes rationality, or acting logically, as the indicator for success. Both of these approaches work from the inside out and assume there is a internal representation that generates human-like behavior. So far both have failed in capturing the humanness described previously.
My propositions are in two parts. First, phenomenological agents may be better accepted as interactive agents, because they are more predictable and thus more comfortable to the user during encounters. Second, techniques for modeling phenomenological agents should be applicable in modeling certain aspects of humans, so it may be possible to have a unifying view of the agent, the human, and the world environment (all of which are phenomenological) that can be computed.
One obvious precursor to phenomenological modeling is statistical modeling. In natural language processing, the corpus-based approach is a statistical process that learns from a large collection of language samples. Learning from a language corpus starts from an a priori statistical model such as the hidden Markov model (HMM). After training, the HMM accepts and generates text that are statistically similar to the training samples.
Another approach that led to phenomenological modeling is Frege's Principle, due to Gottlob Frege of symbolic logic. Following Frege's Principle, a model for airline would be a listing containing all instances of airlines in the world. Similiarly, a model for the concept of blue would consist of all blue items in the world.
Phenomenological modeling captures the behavior of an agent. The agent can be a software agent, a robot, a person, or the world. It can also be a language, or a concept set under Frege's Principle.
In phenomenological modeling, tracing to its origins in philosophy, we do not assume any underlying structure that might govern the generation of the observed patterns. Nothing is material in phenomenology besides the recurring patterns themselves.
The fundamental assumption of phenomenological modeling has that the observable instances of an agent constitute the entirety of the agent. The degree of fidelity of a phenomenological model is the extent of observed instances the model encodes.
Modeling isolated incidents alone is often insufficient in capturing acts of consciousness. Temporal sequences that are vital to human cognition is lost when viewed in isolated time slices. This is especially true in natural languages where preceeding words give the context for interpreting the words following.
Philosophy has established useful results for us. As Husserl puts it, "consciousness must reach out beyond the now." Time consciousness, in Husserl's terms, has three components: primal impression (now), retention (past), and protention (future). Modeling intelligent acts therefore would require recording the sequence consisting of the three components.
[...]
Last Update: February 3, 1997
Created: February 1, 1997