Cognitive Ontological Person Experiment COPE

Concept and Goal: Research in the area of cognitive computing covers a wide range of approaches from studying how birds sing to building robots. This proposed project provides an opportunity to explore the concept of creating a semantic simulation of some of the important decision making characteristics of humans. Important, in this context, means those things about a human that would result in a difference in behavior based on some stimuli. Since almost every human characteristic probably has some influence on behavior for some particular stimulus it is necessary to limit consideration of the types of stimuli in order to somewhat bound the number of characteristics to be represented. The ultimate goal of the project is to create a simulation that will respond to stimuli in exactly the same manner that the human who is modeled in the simulation would react. This provides for a very easy testing environment. The human can be provided a set of stimuli and the resulting reactions can be recorded. The same stimuli are then provided to the simulation and the reactions again recorded. The two results are then compared. If there are no difference in the results (you can't tell the difference between the human reaction and simulation reaction: a simple Turing Test), the experiment is considered successful. The process begins with very simple stimuli and progressively moves to more complex ranges of stimuli as the complexity and correctness of the simulation evolves. Correctness, in this sense, does not necessarily reflect how well the simulation models specific human cognitive processes, but only how well the simulation results compare to the human's performance. That is not to say  that theories drawn from the study of how animals think, including humans, are not important and may not provide key insights into the best simulation approaches, but only that the results are what matters. If we can duplicate the human's reactions across some reasonable set of stimuli, the contention is that the simulation will have demonstrated some level of cognition. The breath of the stimuli that is reasonable is arbitrary, but it is likely that the threshold will be at a point when the simulation performs "useful" human-equivalent decisions. In the context of this experiment, cognition and decision making is the principle area of investigation. Therefore, the representation and modeling of physical characteristics is only important when those characteristics would directly impact the decision making process. For example, the fact that a person has a severe physical handicap really isn't a major factor on how that person might decide to vote on an issue, unless the issue is related to providing better facilities for handicapped people, in which case the fact of being handicapped might influence the person's decision process.

Success Criteria: To expand on the concept of usefulness of the COPE simulation consider some of the potential applications discussed below:
- Productivity Enhancement: It would be very useful to have a 24/7 version of oneself active on the Web making decisions, if there was a reasonable level of confidence the resulting decision were in fact the same ones you would have made given the same stimuli. This would be true even if the range of stimuli was somewhat limited, as long as the decision making tasks were things that benefited you. For example, if you wanted to buy a car, and if the simulation reacted the same as you would to information about buying a new car, you might find it very useful to have the simulation evaluate a very large set of information, including pricing and payment terms, from dealers and individuals wanting to sell you a car. You might, or might not, empower the simulation to complete the financial transaction for something as expensive as a car, but if the simulation was evaluating the information for the purchase of a new CD you wanted you might go ahead and empower it to make the purchase. An online shadow-self might do many other useful things related to business decision, analysis of information, etc.
- Life After Death: This may seem a little strange at first, but it might be very desirable for some people to have a representation of themselves "living" on the Web long after they have physically died. Imagine a family tree where you could click on a relative's picture and link to a realistic simulation of that person. Your great-great-great grandchild might link to your COPE persona and ask it questions about how you would have felt and responded to various situations. If there was a high degree of confidence the simulation would respond as you would have in person, that might be a very rewarding experience for your future descendants. Such a concept might even extend to capturing the decision process for a great thinker so that future generation could ask that person's avatar how it would respond to situation far in the future.
- Criminal Behavior: Might it be useful to capture the criminal "mind" of violent criminals while there are incarcerated? There are certainty privacy issues here, but the value to society might overcome those concerns in some circumstances. For example, capturing the decision process for a child molester might allow law enforcement or parole boards to explore how that person would react to certain stimuli in the future. Of course, capturing and measuring the decision making process for someone who may not be fully cooperative would be much more difficult, but subtle psychological testing designed to extract the needed decision processing information might still provide a very accurate simulation.
- Enemies or Adversaries: This is similar to the concept of simulating specific criminal behavior above, but the privacy issues are probably of less concern. In this application you would model potential enemies or adversaries in a attempt to evaluate how they would respond to various stimuli. However, since these people will be much less likely to provide the level of detail needed to parameterize the simulation, the degree of accuracy of response would be less certain. Never-the-less, the lessons and experience learned in creating simulation of cooperating subjects should be helpful in deciding the most important parameters for non-cooperative subjects.
- Medical Treatment: If a person's key mental characteristics were represented accurately enough, a psychologist might be able to experiment with more radical treatments than he/she would risk using directly on a human.

Constraints: This is a very ambitions project - possibly not even achievable. It could even be characterized as a Grand Challenge type problem. Therefore, to have any reasonable chance of success certain constraints must be imposed. Some of these are discussed below:

Input/Output: The means of communication to and from the semantic simulation should not be a major concern or resource driver. It really doesn't matter if the simulation responds to stimuli in perfect English, French, Russian, Spanish, sign language, or Morris Code. The important issue is that its responses are the same as the human it is modeling. It is important that all input and output be via a web interface. This will make it much simpler for evaluators and peers to test various ontology and agent combinations and to examine and modify different ontology representations.

Model-Driven Approach: There are many schools of though about the correct, or even best, way to achieve computer cognition. This approach arbitrarily selects a model-driven approach in order to bound the problem and further explore this approach. Such an experiment has benefits to the entire field of cognitive research in that after some reasonable level of effort progress has not been shown it may help to narrow the search to alternative approaches.

Semantic Web: The Semantic Web is targeted as the environment for hosting this project, so the representation languages will be the Resource Description Language - RDF, Web Ontology Language - OWL and possibly the Semantic Web Rules Language -SWRL.  The premise in not that the Semantic Web provides any special functionality that would ensure success, although it might. However, the Semantic Web does provide some distinct advantageous for such a project. It is based on open World Wide Web Consortium (W3C) standards that will ensure the interoperability of trial representations. It provides a distributed development, demonstration and testing environment that will foster the full exchange of ideas. Each researcher can easily maintain their own versions of representations and conduct independent experiments while making their results public at the same time. The descriptive logic and rules-based representation languages (RDF, OWL, SWRL) are expressive enough to allow for a wide range of representation. Doug Lenat of CyCorp has stated that 95 to 98 percent of the Cyc knowledge base could be represented using OWL. Semantic Web Services expressed using the Semantic Web Services Language, OWL-S, can be used as desired.

Agent Languages:  Software agents written to process the human characteristics represented within the Semantic Web-based ontologies can be written in any computer language desired by the researchers. The only criteria is that the agents must interact with the Semantic representations written in RDF, OWL or SWRL and with the web interface.

Architecture: The basic COPE architecture concept is illustrated below:


Cognitive Ontological Person Experiment (COPE) Architecture

Web-Based Ontologies:  The ontologies will represent human characteristics that are relevant to human decision making and responses to stimuli. The initial characteristics will be relatively simple and represent core human behavior drivers, such as love, hate, loyalty, and fear. Over time, and with experimental success, the representations will become much more complex.

FOAF:  The friend of a friend (FOAF) initiative may provides a reasonable starting point for experimentation. FOAF uses Semantic Web technologies and already has a large number of Semantic Web researchers contributing to its development. FOAF captures some basic information about a person such as name, email address, home and work web sites and the property "knows". The knows property provides a means to identify other people with whom one is acquainted. Relationships, a vocabulary for describing the relationships between people, is an extension to FOAF that quantifies the knows property. It provides subproperties such as: parentOf, childOf, closeFriendOf, knowsByReputation. There are currently twenty-five of these quantifying properties. A jpeg version of the VisioOWL graphical representation of the FOAF and Relationship ontologies can be seen below. 


Emotion Ontology: This gives us a good start on representing a person, either real or simulated, but would only allow a software agent to elicit realistic responses to simple questions about the person and his/her acquaintances. The eventual goal is to incrementally add characteristics to the COPE simulation to the point where the simulated person possesses all of the key characteristics of a real person. Probably the best approach is to start with a few well-defined characteristics and continue to add new characteristics centered around simple concepts.

A first step might be to further extend the ontology to describe how the simulated person "feels" about the people already represented in the FOAF and Relationships ontologies. Some basic emotions describing how the simulated  person feels about people are: love, hate, envy, disdain, jealousy, contempt, pride, shame, etc. There is a subtle difference in describing how you feel about other people and how you feel about yourself. You may feel happy, sad, joyous, angry, etc. independent of specific people. For example, you might feel angry about getting fired but not angry at the person who fired you because you understand that person had no choice in firing you, possibly because the company is losing money and you are the person with the least longevity. As a first extension we will try to capture emotions related how you feel about people, including how you feel about yourself at any point in time. For example, the fact that Lisa Flynn is my wife makes me feel very happy, but for other undefined reasons I may feel sad overall. The emotions I have listed are arbitrarily based on a cursory search of the web and will be refined and extended through dialog as this exercise progresses. The VisioOWL graphical representation shows the addition of these emotions as related to the Person Class in the FOAF ontology is shown below.


Concept for Person Ontology that Extends FOAF to Include Additional Concepts Such As Emotions, Beliefs, Goals, Standards, and Standards

. Note that the current set of emotions are sets of opposites, such as love/hate. It is reasonable to consider degrees of love or hate, so I have arbitrarily chosen the value +10 to represent the maximum amount of love and the value -10 for the maximum degree of hate. A web interface might be used to make it easy to set and adjust the degree of each emotion with respect to each of the people represented in the FOAF ontology. An simple example of such a web interface is shown below.

.


Agents: Since the ontologies representing human characteristics use a descriptive logic language, OWL, there is no procedural processing capabilities inherent in the ontologies themselves. Software agents are used to code algorithms to process the information contained in the ontologies and to set and modify instance values of ontological classes and properties.

Experimental Process: An iterative process for progressively complex experiments is recommended to lead to cognitive success.