Architectural design of an emotional agent with personality by



Download 96.69 Kb.
Date conversion03.05.2017
Size96.69 Kb.


ARCHITECTURAL DESIGN OF AN EMOTIONAL AGENT WITH PERSONALITY

By

Sreekanth S. Nemani


A report submitted in partial fulfillment

of the requirements for the degree


of

MASTER OF SCIENCE

in

Computer Science


Approved:
______________________________ _______________________________

Dr. Vicki H. Allan Dr. Stephen J. Allan

Major Professor Committee Member

______________________________ _______________________________

Dr. Gregory W. Jones Dr. Donald H. Cooley

Committee Member Department Head


UTAH STATE UNIVERSITY

Logan, Utah

2003
ACKNOWLEDGMENTS


I would like to thank my committee chairperson Dr.Vicki Allan for her guidance and assistance in writing this report. I would also like to thank her for her support in the research that led to this report. I would like to thank Dr. Stephen J. Allan and Dr. Gregory W. Jones for being on my report committee.

I would like to give special thanks to my colleagues Mr.Chaitanya Gharpure, Mr.Kiran Badam and Mr.Vikramaditya Sri Vatsavaya for their assistance in the research leading to this report. I would also like to thank my parents and friends for their patience, encouragement and moral support throughout the process of preparing this report.

Sreekanth S. Nemani


CONTENTS

Page


ACKNOWLEDGMENTS…………………………………………………………………… ii

LIST OF FIGURES………………………………………………………………………….. iv

ABSTRACT………………………………………………………………………………....... v

CHAPTER

1 INTRODUCTION……………………………………………………… ………….. 1

1.1 JADE………………………………………………………………... 2

1.2 OpenCyc…………………………………………………………….. 3
2 AGENTS AND THE ALGEBRA OF EMOTIONS………………………………. 6

2.1 Introduction…………………………………………………………. 7

2.2 Negotiating Agents………………………………………………….. 9

2.3 Predicting Marital Stability…………………………………………. 10

2.4 Emotions…………………………………………………………….. 13

2.4.1 The Transformations……………………………………….. 17

2.4.2 The Process………………………………………………… 19

2.5 An Example: Movie Plan…………………………………………… 20

2.5.1 Emotional State Representation……………………………. 20

2.5.2 Emotional Transformations………………………………... 23

2.5.3 Response…………………………………………………… 27

2.6 Conclusions………………………………………………………… 28

2.7 References………………………………………………………….. 28
3 MATES: A CONVERSATIONAL AGENT SYSTEM…………………………... 32

3.1 Introduction………………………………………………………… 32

3.2 Previous Work…………………………………………………….... 35

3.3 MATES Agent Architecture………………………………………... 37

3.3.1 Architectural Details………………………………………. 37

3.3.2 Three Layered Architecture……………………………….. 39

3.3.3 Survey……………………………………………………... 41

3.4 Results…………………………………………………………….... 41

3.5 Conclusions……………………………………………………….... 43

3.6 Future Work……………………………………………………….. 44

3.7 Acknowledgments………………………………………………..... 44

3.8 References………………………………………………………… . 44


4 SUMMARY AND CONCLUSIONS……………………………………………... 49 REFERENCES……………………………………………………………………………... 51

LIST OF FIGURES
Figure Page

  1. Conversation Plan………………………………………………………………… 10

  2. Matrix representation of the Emotional State of an Agent………………………... 14

  3. The MATES Agent Architecture………………………………………………….. 38


ABSTRACT

Architectural Design of an Emotional Agent

with Personality

by

Sreekanth S. Nemani, Master of Science



Utah State University, 2003

Major Professor: Dr. Vicki H. Allan

Department: Computer Science

Many emotional and social agents exist that model various aspects of human behavior and personality. This report presents the architectural details of an emotional agent called MATES (Marital Agent Trait-based Emotion System) that is currently under development. The MATES agent is an intelligent agent with personality, emotions, goals, and planning that converses with another MATES agent modeled as its spouse.

The report delves into the model of emotions developed for the NAMS agent, a preliminary version of the MATES agent. The Architectural design of the MATES agent is described in detail outlining the various components of the system. A MATES agent is unique in the way it handles planning, emotions and the inference engine. It uses features from various existing concepts like BDI (Belief, Desire, and Intention) architectures, three layer mobile robot architectures etc., and combines them with some new concepts like high-level planning.

The MATES agent is envisioned as a useful tool for relationship counselors, married and unmarried couples, and agent developers.

(53 pages)

CHAPTER 1

INTRODUCTION

There is no standard definition for Intelligent Agents. The dictionary defines an agent as one that has the authority to act for or represent another. An intelligent agent can be considered as a persistent, autonomous entity that acts on behalf of another. An agent realizes a specific set of goals for which it was designed. Weiss[20] defines an intelligent agent as “one that is capable of flexible autonomous action in order to meet its design objectives, where flexibility means: Reactivity, Pro-activeness and Social-ability”

Agents are used in many applications involving large-scale distributed applications like Cougaar[5]. Agents have been used in distributed network management[6] and dynamic routing[12]. Agents have also been used in developing business negotiation applications like e-commerce[13][14]. Intelligent agents are specifically being used extensively in applications involving distributed artificial intelligence like modeling humans and human interactions[3][19].

In this research we have attempted to develop agents that model the personality, emotions and behavior of an existing person. Two such agents then converse with each other on a variety of issues. The agents do not model a complete human as it is still beyond the scope of present psychological research but instead attempt to produce a believable conversation. The domain we have chosen to work is couple relationships involving emotions, personality, cognition, planning, goals etc…

In this scenario two agents are programmed, one with the personality information obtained from the husband and the other with the personality information from the wife. These two agents then converse on relevant topics based on their goals and personality. The conversation is generated dynamically as each agent receives the response from the other agent.

Our intelligent agent is called MATES (Marital Agent Trait-based Emotion System). A MATES agent has a personality which is programmed by the user, an emotion system that generates the appropriate emotions depending on the scenario, goal processing capability to try and achieve its goals, planning capability, reactive behavior and a knowledge base.

This tool can be used by couples and counselors to analyze the personality traits and their effect on the conversation and the relationship[10][17]. It can also be used to predict the future of a relationship based on the personality and the past interactions[4][8][9]. The change in relationships over time can also be studied.

The prototype MATES agent is still under development and its goal is to demonstrate the role of personality attributes in negotiation. A couple who is considering marriage (whom we designate as Bob and Alice) use the MATES agents in the following way. One MATES agent is programmed with the personality data from Alice, while another MATES agent is programmed with the personality data from Bob. The two MATES agents can then interact on any issue generating a believable conversation. This report gives the innovative design details of various components of the MATES prototype system.

The MATES agent uses JADE (Java Agent Development Environment)[1][2] as the multi-agent distributed environment and OpenCyc[16] as its knowledgebase.


    1. JADE

There are many Multi-agent systems that can be used to develop agents like JATLite[11], RETSINA[18], JADE[1][2], etc.

We have chosen JADE as our agent development platform. JADE[1][2] is a software development framework aimed at developing FIPA[7] (Foundation for Intelligent Physical Agents) compliant multi-agents systems. JADE agents are distributed and rest on the FIPA compliant JADE platform. Every system in the agent network has one JADE agent platform. Each JADE platform can have many JADE agents.

JADE agents communicate with each other using a FIPA compliant Agent Communication Language (ACL). The full FIPA communication model using interaction protocols, ACL, content languages, envelope, encoding schemes, ontology and transport protocols have been implemented in JADE. A JADE agent uses Java RMI, event notification and IIOP to communicate with other agents. It chooses the appropriate mode of communication based on the location of the other agent.

We have chosen JADE as our agent platform for the following reasons:



  1. It is a FIPA[7] (Foundation for Intelligent Physical Agents) compliant agent platform.

  2. The agent framework is developed in Java and provides a very extensive set of agent classes and utilities.

  3. We can develop our agents using Java and the JADE framework and hence the learning curve was small.

  4. It provides a FIPA compliant Agent Communication Language (ACL).

The JADE agent is essentially a shell having communication capabilities, a distributed platform, specific language constructs for message content, and everything needed for autonomic behavior. A JADE agent is not intelligent, only autonomous and distributed. The MATES agent uses JADE as a shell and its own architecture for intelligent behavior.




    1. OpenCyc

OpenCyc[16] is a general Knowledge base and reasoning engine. It is an open source version of the Cyc technology developed by Cycorp, Inc. It has a reasoning engine that can be used to code predicate logic, a planner and a huge database of general facts. It consists of 6000 concepts and 60,000 assertions about those concepts forming the common sense knowledge of OpenCyc. It operates as a server and provides a socket interface.

OpenCyc uses CycL as the language for knowledge representation. SubL is a language that forms a layer on top of CycL for complex querying and inference. SubL is similar to LISP in its working and construction, and uses the predicates that have been developed in CycL. The OpenCyc knowledge base can be interfaced with the Java using the Java API provided by it. We use this Java API to connect the knowledge base to our agent developed in JADE.

The MATES agent uses OpenCyc to develop its knowledge base and intelligence. It then interfaces this knowledge base with the agent shell developed in JADE to create a complete autonomous, distributed and intelligent agent.

While major portions of the MATES agent have been developed allowing us to test the various approaches adopted, the complete MATES system is still under development and what we have is just a prototypical agent.

I have been involved with this project from the beginning and have been one of its major contributors. My contribution to the project has been at various levels:


  1. I performed the initial literature survey for the research, and was instrumental in deciding the direction of research and the selection of the necessary platforms and tools.

  2. I am the primary author and the major contributor for two technical papers (given in chapters 2 and 3).

  3. I designed the architecture of both the NAMS and the MATES agent systems.

  4. I developed the emotion model that was used in the NAMS agent and a simplified version of which was used in the MATES agent.

  5. I was the sole developer of the NAMS agent – designing the structure of the knowledgebase, developing the inference engine and the emotion system in prolog.

In addition, I also partly contributed to:



  • Porting the inference engine on to OpenCyc.

  • Deciding on the working of the planner.

  • The integration of the surveys into the overall system.

This report is organized in the article format. Chapter 2 is an article titled “Agents and the Algebra of Emotion” presented at the international joint conference on Autonomous Agents and the Multi-agent systems (AAMAS), Melbourne, Australia, 2003. This chapter describes the model of emotions that was used in the development of the NAMS (Negotiating Agents and Marital Stability) agent, an initial version of the MATES agent. Chapter 3 is an article titled “MATES: a conversational agent system” that shall shortly be submitted to a conference. Finally chapter 4 gives the conclusions.


CHAPTER 2 (see separate document)

CHAPTER 3

MATES: A Conversational Agent System

S.S. Nemani

Department of Computer Science, Utah State University

Logan, UT 84322-4205, E-mail: ssnemani@cc.usu.edu

V. H. Allan

Department of Computer Science, Utah State University

Logan, UT 84322-4205, E-mail: allanv@cs.usu.edu
Couple interaction is a complex issue as it involves volatile human behavior, personality, and emotions. Using MATES (Marital Agent Trait-based Emotion System), we program two individual agents to take on the personality of a couple (Alice and Bob). These agents then converse with each other based on the current plans, goals, events, history, personality, and emotions.

This paper describes the basic features that the MATES agent requires and the architecture that we have developed to support them. It uses a layered approach using surveys to parameterize a goal database, a high level dynamic planner, and a complex emotion system.
1. INTRODUCTION

The MATES agent is envisioned as a tool used by pre-marital couples to understand conflicts and issues. It will be used to help couples realize the origins of arguments, personality tendencies and their effect on the harmony of their relationship. It could also be used by marriage counselors and researchers to study human interaction in different situations. We bring interpersonal activities to the virtual stage by encoding agents with the emotions and traits of the premarital couple they represent. While a completely realistic system is beyond the scope of current psychological understanding, a system which is believable is within reach [26]. Marital stability is used as an application area to test our hypotheses for the following reasons: (a) couple relationships are emotionally charged, (b) there is a rich body of literature in the area of couple interaction, (c) there is a well established network of professionals and participants to evaluate the system, (d) accuracy of reporting is increased when two individuals report. A key motivation underlying the proposed research is that the costs of modeling and evaluating relationships through simulation before a commitment is made are far smaller than counseling real participants after poor relationship decisions have actually been made. This assumption reiterates the assumptions of the Computational Organization Theory[5].

Consider a couple, Alice and Bob. One MATES agent is programmed with the personality information of Alice while another MATES agent is programmed with the personality information from Bob. The personality data is obtained by using IPIP-NEO surveys completed by Alice and Bob[9][10][18]. The two MATES agents converse with each other about a pre-selected topic and the course of the conversation varies depending on their personality, goals, and interaction style. We can also adjust the personality of these agents thus enabling us to answer what if kinds of questions.

Some of the unique features that a MATES agent incorporates are:



  1. Dynamic behavior – There is no reasonable methodology to predict the response(s) an agent receives from another agent. Thus, dynamic re-planning is an essential part of a MATES agent.




  1. Emotions – Humans are emotional beings and hence a MATES agent incorporates emotions in order to be believable. Emotions, in fact, play a major role in conversation. MATES uses the fundamental emotions as specified by Izard[19]. The MATES agent uses an algebraic representation of emotions, based on that developed for the initial version, the NAMS agent[21]. In this representation, the emotional state of an agent is represented in the form of a vector, where the elements are the intensities of the eleven fundamental emotions.




  1. Emotional transformations - There are many different factors that can dynamically affect the emotions motivating a transformation of emotions. Gmytrasiewicz[8] gives a mathematical representation transformations. The MATES agent builds on the emotion transformations developed by us for the NAMS agent[21]. The five transformations we use in a MATES agent are: Addition, Combination, Degradation, Attribution, and Comparison.




  1. Personality modulation – Various personalities respond in different ways to similar situations. As we program the agents, we are forced to look at emotions and couple interactions in ways not previously clear. The primary goal of the MATES system is to enable couples, researchers, and counselors to observe and analyze the conversation as it predicts marital stability. It helps a couple consider appropriate changes by answering questions like “What if I were more agreeable?” or “What if we were more honest?.” The personality of a MATES agent can be modulated appropriately to observe the difference such interaction changes make to the conversation.




  1. Conversation predisposing – A variety of conversational paths are possible. The selection process is facilitated using a modified version of Gottmann research [11]. Gottmann is well known for his work in analysis that is able to predict marital stability. We examine the issues Gottmann identifies in therapist interactions, but attempt to extract these through self-reporting via a relationship questionnaire [11][12]. This questionnaire identifies goals, interaction style (self-soothing, repair attempts, criticism, contempt, stonewalling), and relationship stressors. Collecting this information allows us to predispose the conversation to end in conflict or be resolved peaceably. While the personality survey concentrates on individual attributes, the relationship questionnaire is concerned with the interactions of the couple.




  1. Goals – A MATES agent has goals to reach. The goals of a MATES agent can be selected from a database of goals common to many pre-marital couples. The relationship questionnaire is used to select relevant goals and attribute appropriate emotions to their progress. A MATES agent is not just reactive in nature but is also proactive in trying to achieve its goals. Hence, even as the agent reacts to an event based on its emotions and personality, it also constantly strives to achieve its goals.


2. PREVIOUS WORK

The Emile system developed by Gratch[13][14][15] is an emotional agent system which depends heavily on planning. It uses its plan to determine the actions that will further its goals, and relates the events that occur to its plan. It then uses a plan-based appraisal where emotions are generated based on how the events have affected the plan. An equation based on Velasquez’s cathexis equation[28] to calculate the intensity of these emotions is applied. Emile’ is not designed to deal with dynamic re-planning. Its plans are very low-level in nature, thus rendering it practically impossible to perform any kind of dynamic re-planning. It also does not allow us to modify the personality to observe different behaviors.

In the AlphaWolves[27] system, Tomlinson et.al. have developed a unique system to partially control their animated wolves. The user can direct the actions of the wolves, while the system still autonomously generates and maintains emotions. They have thus preserved the realistic behavior, but at the same time their system does not involve any planning. The AlphaWolves system is reactive in nature when the wolves are not controlled by the users. When it is controlled by the user, the AlphaWolves system only involves itself in executing a selected action based on personality and emotion.

On the other hand, the SCREAM system developed by Helmut and Mitsuru[24][25] uses scripting to allow the users to modify the system to their domain. It basically deals with emotion processing and does not deal with planning and goal achievement.

Marcus Huber’s JAM[17] system is based on the popular BDI architecture suggested by Bratman et.al.[1] and gives the general architecture of an agent and how each of its components interact with each other. It deals with plans, goals, desires and intentions but completely ignores emotions. Like Emile, its planning is more low level in nature.

Other systems like Reilly’s[26] social agents have used emotion and personality in negotiation, while Camuri and Coglio’s[4] emotional agents deal with the interdependence of the various components of an emotion model. Marcella and Gratch[16] have used coping as a technique to dynamically modify beliefs and goals and generate irrationality in behavior. Bazzan and Bordini[3] have used the OCC theory by Ortony et.al.[23] to create Emotion Groups and a set of mathematical rules to generate each of the emotions.

In the field of robotics, there is a set of very interesting models called the three-layer architectures. Most of the three layer architectures consist of a reactive feedback mechanism, a reactive plan execution mechanism and a mechanism for deliberative computation[7]. In Gat’s ATLANTIS[6], these layers are called the controller, the sequencer, and the deliberator[7]. These architectures are specifically developed for the mobile robots and do not include many crucial aspects necessary for a MATES agent, like emotions and dynamic re-planning.


3. MATES AGENT ARCHITECTURE

3.1 ARCHITECTURAL DETAILS

Figure 1 is the architectural diagram of a MATES agent. The MATES agent has a knowledge base consisting of Beliefs, History, Personality, and the Goal database. It also has three processing layers: the Plan generator, the Inference engine, and the Reactive response generator. These layers use the information from the knowledge base to process the data appropriately.

The Beliefs component stores the likes/dislikes of the agent. This is general information: its perception of the likes/dislikes of the partner agent, its perception of the goals and personality of the other agent, its own goals and probability of success in achieving the goals. Whenever an external communication is received, it is analyzed for content, and it is added to the beliefs store. Some sample beliefs in the knowledge base could be:

I love apples

I find animals disgusting

There is a party this Friday

Alice likes chick-flicks.

Alice wants a stable relationship
The Personality component is a user defined component. We follow the Big Five model of personality[2][20]. Personality in this model is measured by the five factors: Extraversion, Openness, Agreeableness, Conscientiousness, and Neuroticism. To obtain the personality values of a human, we use the IPIP-NEO questionnaires. The IPIP-NEO is shorter version based on the commercial inventory, the NEO PI-R™, authored by Paul T. Costa, Jr. and Robert R. McCrae. The original survey is considered by many psychologists to be the best inventory for measuring traits within the Five Factor Model (FFM) of personality [9][10][18]. The questionnaire surveys each of the five factors of personality in multiple ways. The bigger the questionnaire the more accurate the personality values that we derive. There are the 50, 100, and 300 question IPIP surveys that can be used. The Personality directly influences all the three processing layers of a MATES agent.

The History component stores the history of the conversation – the emotions, actions and events that occur in a chronological order.





Figure 1 The MATES Agent Architecture


The Goal Database consists of high-level goals that an agent can work to achieve by acting in a specific way on a pre-existing state. This database is developed using a survey as described in the section 3.3. There are many such high-level goals in the goal database. The plan generator looks at the final goal to be achieved and strings together a set of high-level goals to form a plan.

The Inference Engine consists of the emotion transformer and the behavior generator. These two together perform the low level handling of a goal. When the inference engine receives the plan from the plan generator, it examines the first goal in the sequence to be achieved. The emotion transformer changes the present emotional state based on the plan and the personality. The behavior generator considers the plan, the event, the personality and the emotions to generate a corresponding behavior.


The Reactive Response Generator creates a response to be communicated based on the behavior generated and the emotions. This component attempts to generate a reflex response when it initially receives a communication and if it fails it creates a wrapper to the response generated by the inference engine.
3.2 THREE LAYERED ARCHITECTURE

There are three basic layers to this architecture. The plan generator is at the top, the inference engine is the middle layer, and the reactive response generator is the lowest layer.

Any communication received from the conversation partner is received by the reactive response generator. It checks if there is any reflex action to be performed (e.g., a question requesting further information) and performs it. If there is not any reflex reaction, then the control goes to the inference engine for further processing. The inference engine attempts to take action based on the current plan, emotional state, and personality. If the inference engine fails to produce a response then the control goes to the plan generator. The plan generator creates a plan to deal with the situation and sends it down to the inference engine, which then acts based on the plan and sends a response down to be communicated to the other agent.

The plan generator uses high-level goals from the goal database to create a plan, and each of these high-level goals are processed and achieved by the middle layer inference engine. The inference engine looks at the plan and tries to achieve the current high-level goal in the sequence. It then considers all the possible ways the high-level goal can be achieved and selects the option that best suits the agent’s personality. OpenCyc[22] allows the relevant choices to be identified and established in a micro-theory, thus making selection efficient as inappropriate choices for the given agent are not present in the data base.

Once the inference engine generates an initial behavior and emotion, the reactive response generator wraps the behavior and emotion in a proper response and sends it to the other agent. Depending on the response from the partner MATES agent, this agent might then decide to react reactively, or continue with the next stage of the plan, or dynamically modify its plan to suit the changed scenario, or ask the inference engine to find other ways to achieve the same plan.

Since the plan generator only deals with high-level goals, dynamically changing plans is possible without overloading the system. Though the fine details are not considered by the planner, the system works since the inference engine considers all possible options to achieving the goal.

This layered approach gives us the capability to dynamically re-plan in an efficient manner. The plan generator only deals with high-level goals to create its plan, and does not involve itself in the nitty-gritty details of a goal. Also, high-level planning is more intuitive in nature as it replicates the human cognitive behavior. Humans plan at a more general level and often omit the specific details. For example, a person might plan to go to the movie tonight, but he/she may not plan in advance every inquiry and response. Instead, he/she has a set of major steps that he/she figures can be achieved and gets them done one after the other, adapting to changing circumstances. Similarly in a MATES agent, a plan is an ordered sequence of high-level goals to be achieved. The Inference engine achieves each goal based on the agent’s personality.

For example, the MATES agent Bob might have a goal of ‘havingFunTonight’, for which the Plan generator creates a plan:



  1. gaugePartnerMood

  2. findPartnerInterests

  3. planMutualFunActivity

  4. scheduleTimingAndTransport

The inference engine attempts to achieve the first goal – ‘gaugePartnerMood’. It might attempt to do this in a variety of ways (depending on personality, history, and interaction style): ask, look for signs, make idle conversation, etc. The appropriate behavior and emotions are generated and communicated to Alice through the response generator. Depending on Alice’s response, the plan generator decides what to do next.


3.3 SURVEY

The goal database is the most critical component in our architecture. The goals in the goal database need to be high-level goals which are reasonably achieved by the inference engine. A proper goal database is a necessity for the functioning of our architecture. If the goals are too low-level in nature, then the plan generator is heavily overloaded, and we have system delay. If the goals are too high level, then the inference engine might not be able to perform to our satisfaction.

Once we have a goal database established, we then have to decide which of these goals are active or relevant to a particular agent. There are goals which are common to all humans, and are persistent through life for example ‘beLiked’. There are others which are transient and are also dependent upon personality. The relationship questionnaire is an ongoing effort with the goal of developing a simple system that can believably direct the MATES agent.
4. RESULTS
The preliminary version of the MATES agent has been developed using the OpenCyc[22] knowledge base and reasoning engine. The conversation is shown as taking place in English, but it is in fact taking place with conversation structures, the emotions communicated as vectors.
Alice: ‘Would you like to go to a movie tonight?’
Bob considers this communication from Alice and since the information is incomplete he responds reactively asking for more details.
Bob: ‘What movie?’
And Alice responds,

Alice: ‘Charlie’s Angels’.
Bob considers this communication from Alice and since he has a presently active goal of ‘haveFunTonight’, he creates a plan dynamically:


  1. decideFunActivity

  2. finalizeActivity

The Inference engine then performs actions to achieve ‘decideFunActivity’ based on Bob’s personality. Bob looks at his likes/dislikes and realizes that he does not like Charlie’s Angels. So the inference engine responds appropriately.


Bob: ‘I am not wild about chick flicks. How about X-Men 2?’
And Alice responds,
Alice: ‘I have never heard of it’.
Upon receiving this response, Bob realizes from his beliefs of Alice that Alice wouldn’t like an action movie. His goal ‘keepPartnerHappy’ is threatened. So, the Plan generator considering the changed priorities, changes the plan to – ‘doWhatPartnerLikes’.

Bob: ‘We could see Charlie’s Angels if you would really like to’.
And Alice responds,
Alice: ‘Okay. That would be great. I like you.’
Bob receives this message and his natural reflex is to respond to compliments with a smile (a non-verbal communication).
Bob:
As we can see, high-level planning helps us dynamically re-plan quickly and conveniently. Also, a simple inference engine can take care of selecting the right response to achieve a goal.
5. CONCLUSIONS

An important aspect of the architecture is the layered approach which helps dynamically re-plan efficiently. Also, this helps model human cognitive thinking. Similar to human cognitive behavior, the MATES planner works at a general level and never plans all the details.


An important facet of this research is that we are attempting to model real people. A MATES agent ‘Bob’ is a model of a real person. The two questionnaires we have used are another important aspect of our architecture as it allows us to define the active goals of an agent while using established methods.

Also, personality influences decisions at every point in the architecture. We have a system for modulating personality in our MATES agent since a change in the personality propagates to the behavior of the MATES agent.


6. FUTURE WORK

We are working to extend our prototype data base to be richer. Work is also progressing for the development of a survey to determine the goals of the agent. A natural language processor and visual representation of the agents is also being developed.


7. ACKNOWLEDGMENTS

Chaitanya Gharpure provided valuable support for the development of the prototype MATES agent and inference engine. Kiran Badam gave valuable support for the development in OpenCyc and the plan generator. Vikramaditya Sri Vatsavaya provided useful support for the development of the personality component. This work was conducted at the Computer Science Department, Utah State University with funding from Community University Research Initiative (CURI).


8. REFERENCES

[1] M.E.Bratman, D.J.Israel and M.E.Pollack. Plans and Resource-Bounded Practical Reasoning. Computational Intelligence, 4(4):349-355, 1988.


[2] D.M.Buss. Social Adaptation and Five Major factors of personality. In J.S.Wiggins (Ed.) The Five Factor Model Of Personality: Theoretical Perspectives (pp.180-207), New York: Guilford.
[3] A.L.C.Bazzan and R.H.Bordini. A Framework for the Simulation of Agents with Emotion. In Proceedings of the Fifth International Conference on Autonomous Agents, Montreal, Canada, 2001.
[4] A.Camurri and A.Coglio. An Architecture for Emotional Agents. IEEE Multimedia, 5(4):24-33, 1998.
[5] K.M.Carley and L.Gasser. Computational Organization Theory. In collection Multiagent Systems (pp. 199-330), MIT Press, Long, England, 1999.
[6] E.Gat. Integrating planning and reaction in a Heterogeneous asynchronous architecture for controlling mobile Robots. In Proceedings of the Tenth national conference on Artificial Intelligence (AAAI), 1992.
[7] E.Gat. On Three layer Architectures. In D.Kortenkamp, R.P.Bonnasso, and R.Murphy, editors, Artificial Intelligence and Mobile Robots. MIT/AAAI Press, 1997.
[8] P.Gmytrasiewicz and C.Lisetti. Using Decision Theory to formalize Emotions for Multi-Agent System Applications: Preliminary Report. In Second ICMAS-2000 Workshop on Game Theoretic and Decision Theoretic Agents, Boston, Massachusetts, 2000.
[9] L.R.Goldberg (1999). A broad-bandwidth, public domain, personality inventory measuring the lower-level facets of several five-factor models. In I.Mervielde, I.Deary, F.De Fruyt, and F.Ostendorf (Eds.), Personality Psychology in Europe, Vol. 7 (pp. 7-28). Tilburg, The Netherlands: Tilburg University Press.
[10] L.R.Goldberg (in press). The comparitive validity of adult personality inventories: Applications of a consumer testing framework. In S.R.Briggs, J.M.Cheek, and E.M.Donahue (Eds.), Handbook of Adult Personality Inventories.
[11] J.M.Gottman. What Predicts Divorce: The Measures. L.Erlbaum, 1996.
[12] J.M.Gottman, J.Coan, S.Carrere and C. Swanson. Predicting marital happiness and stability from newlywed interactions. In Journal of Marriage and the Family (pp.5-22), vol. 60, 1998.
[13] J.Gratch. Emile: Marshalling passions in training and education. In Proceedings of the fourth International Conference on Autonomous Agents, Barcelona, Catalonia, Spain, 2000.
[14] J.Gratch and S. Marcella. Tears and Fears: Modeling emotions and emotional behaviors in synthetic agents. In Proceedings of the Fifth International Conference on Autonomous Agents, Montreal, Canada, 2001.
[15] J.Gratch and S. Marcella. Modeling the Interplay of Emotions and Plans in Multi-Agent simulations. In Proceedings of the Cognitive Science Society,2001.
[16] J.Gratch and S. Marcella. A Step toward Irrationality: using emotion to change belief. In Proceedings of the First International Joint Conference on Autonomous Agents and Multi-Agent Systems, Bologna, Italy, 2002.

[17] M.J.Huber. JAM: A BDI-theoretic Mobile Agent Architecture. In Proceedings of the Third International Conference on Autonomous Agents (Agents ’99), Seattle, WA, 1999.


[18] International Personality Item Pool (2001). A Scientific Collaboratory for the Development of Advanced Measures of Personality Traits and Other Individual Differences (http://ipip.ori.org/). Internet Web Site.
[19] C.Izard. The Psychology of Emotions. Plenum Press, New York, NY, 1991.
[20] O.P.John. The “Big Five” factor taxonomy: Dimensions of personality in the natural language and in questionnaires. In L.A.Pervin (Ed.), Handbook of Personality: Theory and Research (pp. 66-100), New York: Guilford, 1990.
[21] S.S.Nemani and V.H.Allan. Agents and the Algebra of Emotion. In proceedings of the Second International Joint Conference on Autonomous Agents and Multi-Agent Systems, Melbourne, Australia, 2003.
[22] OpenCyc (2002). CycR Knowledge Base CopyrightC 1995-2002 Cycorp, Inc., Austin, TX, USA. (http://www.opencyc.org/). Internet Web Site
[23] A.Ortony, G.L.Clore and A.Collins. The Cognitive Structure of Emotions, Cambridge University Press, Cambridge, UK, 1988.
[24] H.Prendinger and M.Ishizuka. SCREAM: Scripting Emotion based Agent Minds. In Proceedings of the First International Joint Conference on Autonomous Agents and Multi-Agent Systems, Bologna, Italy, 2002.
[25] H.Prendinger and M.Ishizuka. Social Role awareness in Animated Agents. In Proceedings of the Fifth International Conference on Autonomous Agents, Montreal, Canada, 2001.
[26] W.S.N.Reilly. Believable Social and Emotional Agents. Technical Report CMU-CS-96-138, School of Computer Science, Carnegie Mellon University, 1996.
[27] B.Tomlinson, M.Downie, M.Berlin, J.Gray, D.Lyons, J.Cochran and B.Blumberg. Leashing AlphaWolves: Mixing user direction with Autonomous Emotion in a Pack of Semi-Autonomous Virtual Characters. In Proceedings of the 2002 ACM SIGGRAPH Symposium on Computer Animation,San Antonio, Texas.
[28] J.Velasquez. Modeling Emotions and other motivations in Synthetic agents. In Proceedings of the Fourteenth National Conference on Artificial Intelligenc. 1997.

CHAPTER 4

SUMMARY AND CONCLUSIONS

This report has described an innovative matrix algebra based model for agent emotion. The various transformations that can be produced in the emotions and the affects it can have on the conversation have also been outlined. The emphasis has been on developing a model that is capable of representing complex emotions and is adaptable to various situations.

The report has also given the architectural details of the MATES agent. The MATES agent uses a modified version of the innovative emotion model specified in chapter 2. Its architecture is based on a combination of the BDI (Beliefs, Desires and Intentions) architectures, the three layer mobile robot architectures and our unique high level planning. The MATES agent is still a product under development and as yet only partial results have been obtained. The results it has generated have been highly encouraging.

In the current working MATES agent, the emotion generation system, the goal processing system, the inference engine and the reactive response generator are at an advanced stage of development. The development of the plan generator and the complete structure of our knowledgebase are at an intermediate stage of development. Also, currently under development is a system that uses the survey format to gather personality information and convert it to a meaningful format. The current focus of the project is on the mapping of this personality data to the three processing layers. Using surveys to gather personality data with which to model an agent will result in a more realistic system.

The development of the goal database and the selection of the goals from it using surveys require feedback from people specializing in human psychology. We have interacted with human psychology experts on this matter and the development of the resultant survey is in its preliminary stages.

Any reasonable system that models human beings requires a large knowledgebase of facts. The larger the knowledgebase the more believable and realistic the system can get. For this reason, we have used OpenCyc to develop the MATES knowledgebase.

The MATES agent is different from most intelligent agent systems, in that it attempts to model an existing person for that person. Modeling an actual individual can provide us with valuable insights about the factors affecting human behavior and help us to better model future agent systems.

This report has presented the architectural and design details of a system using many unique concepts in combination with some tried and tested methodologies. From a computer science point of view, the MATES agent system aims to provide us with a dynamic and believable conversation generator. On the other hand, the MATES agent system also aims to be a useful tool to relation counselors and marital couples.



REFERENCES

[1] F.Bellifemine, A.Poggi and G.Rimasa. A FIPA Compliant Agent Framework. In Proceedings of PAAM’99, London, England, 1999.


[2] F.Bellifemine, A.Poggi and G.Rimasa. A FIPA2000 Compliant Agent Development Environment. In Proceedings of the Fifth International Conference on Autonomous Agents, Montreal, Canada, 2001.
[3] A.Camurri and A.Coglio. An Architecture for Emotional Agents. IEEE Multimedia, 5(4):24-33, 1998.
[4] K.M.Carley and L.Gasser. Computational Organization Theory. In collection Multiagent Systems (pp. 199-330), MIT Press, Long, England, 1999.
[5] Cougaar (2000). Congnitive Agent Architecture (Cougaar) Open Source Project. (http://www.cougaar.org/). Internet Web Site
[6] T.C.Du, E.Y.Li and A.P.Chang. Mobile Agents in Distributed Network Management. In Communications of the ACM (Vol.46, Iss.7, pp.127-132), ACM Press, 2003.
[7] FIPA. Foundation for Intelligent Physical Agents. (http://www.fipa.org/). Web Site.
[8] J.M.Gottman. What Predicts Divorce: The Measures. L.Erlbaum, 1996.
[9] J.M.Gottman, J.Coan, S.Carrere and C. Swanson. Predicting marital happiness and stability from newlywed interactions. In Journal of Marriage and the Family (pp.5-22), vol. 60, 1998.

[10] C.E.Izard. The Psychology of Emotions. Plenum Press, New York, NY, 1991.


[11] H.Jeon, C.Petrie, M.R.Cutkosky. JATLite: A Java Agent infrastructure with Message Routing. In IEEE Internet Computing 4(2),March/April 2000.
[12] K.H.Kramer, N.Minar and P.Maes. Tutorial: Mobile Software Agents for Dynamic Routing. In ACM SIGMOBILE Mobile Computing and Communications Overview(Vol.3, Iss.2, pp.12-16), ACM Press, 1999.
[13] D.B.lange and M.Oshima. Seven good reasons for Mobile Agents. In Communications of the ACM (Vol.42, Iss.3,pp.88-89), ACM Press, NY, USA, 1999.
[14] P.Maes, R.H.Guttman and A.G.Moukas. Agents that Buy and Sell. In Communications of the ACM (Vol.42, Iss.3, pp.81-ff), ACM Press, NY, USA, 1999.
[15] J.D.Mayer, H.F.Chabot, and K.M.carlsmith. Conation, Affect , and Cognition in Personality. In Cognitive Science Perspectives on Personality and Emotion. Elsevier Science B.V, Amsterdam, Netherlands, 1997.
[16] OpenCyc (2002). CycR Knowledge Base CopyrightC 1995-2002 Cycorp, Inc., Austin, TX, USA. (http://www.opencyc.org/). Internet Web Site
[17] A.Ortony, G.L.Clore and A.Collins. The Cognitive Structure of Emotions, Cambridge University Press, Cambridge, UK, 1988.
[18] K.Sycara and A.S.Pannu. The RETSINA Multi-agent System: Towards Integrating, Planning, Execution and Information Gathering. In Proceedings of the Second International Conference on Autonomous Agents, May 1998.
[19] W.S.N.Reilly. Believable Social and Emotional Agents. Technical Report CMU-CS-96-138, School of Computer Science, Carnegie Mellon University, 1996.

[20] G.Weiss. Multiagent Systems: A Modern approach to Distributed Artificial Intelligence. MIT Press, Cambridge, Massachusetts, 1999.



 This chapter is a paper co-authored with Dr.Vicki H. Allan, Computer Science Department, Utah State University that shall shortly be submitted to a conference.



The database is protected by copyright ©ininet.org 2016
send message

    Main page