A SCHIZOPHRENIC APPROACH FOR INTELLIGENT
CONVERSATIONAL AGENTS
Jean-Claude Heudin
Imedialab, International Institute of Multimedia Léonard de Vinci, Paris La Défense, France
Keywords: Conversational agent, Intelligent agent, Multi-agent, Schizophrenic model.
Abstract: We present a novel approach for creating intelligent conversational agents based on a “schizophrenic”
model implemented using the EVA (Evolutionary Virtual Agent) nano-agent architecture. The Ms House
experiment developed using this approach is compared with Eliza and the Alice chatterbot.
1 INTRODUCTION
Since the first conversational program called Eliza
developed at the MIT in the 60s (Weizenbaum,
1966), there have been a large number of studies for
designing intelligent agents that could dialog in a
natural way with human users. A major part of this
research focused on various aspects of the problem
such as natural language interaction using syntaxic
or semantic approaches, non-verbal communication
such as emotional expressions, user’s graphical
interfaces and self-animated characters (Cassel,
2000). In the meantime, there have been a growing
number of commercial applications using
conversational agents on the web based on
programming toolkits like AIML (Wallace, 2002).
However, no implementation has made a real
breakthrough since the original Weizenbaum’s Eliza
experiment. The main problem is the believability of
the artificial character. Most implementations show
a simple and straitforward personality instead of a
complex and versatile human-like personnality. A
second important problem is the small amount of
information and knowledge of the agent. Most
implementations are based on a limited set of
predefined answers or behaviors with no learning
capabilities.
In this paper we propose a novel “schizophrenic”
approach which potentially addresses these two
problems. We will focus in this article on the first
problem, that is believability, but we argue that our
approach is also promissing for addressing the
second one, that is learning. In section 2, we make
first a brief overview of the EVA bio-inspired
architecture (Heudin, 2004) and its programming
language called nanoScheme used for developping
the approach and prototype. A more detailled
description of the EVA agent technology could be
found in (Heudin, 2010). In section 3, we describe
the “schizophrenic” approach and its principles.
Section 4 describes an experimental prototype based
on this approach: a virtual psychoanalyst called Ms
House. The qualitative efficiency of this prototype is
then compared with an implementation of Eliza and
the Alice conversational engine (Wallace, 2002). We
conclude by outlining future developments.
2 EVA OVERVIEW
2.1 Complex System Approach
The aim of the EVA project is to provide a software
framework for studying machine intelligence.
Natural intelligence emerges from the huge number
of nonlinear interactions that occur within the brain
architecture. The brain is itself the result of millions
years of co-evolution within the earth environment.
Therefore, simulating natural intelligence using a
classical reductionist approach seems not well
adapted to this challenging goal. Moreover, instead
of trying to strictly reproduce human intelligence as
we now it, our long term goal is rather to create
complementary machine intelligence. While human
intelligence is more efficient in most real-life
situations, thanks to its long history of evolution,
machine intelligence will be far more efficient for
251
Heudin J..
A SCHIZOPHRENIC APPROACH FOR INTELLIGENT CONVERSATIONAL AGENTS.
DOI: 10.5220/0003183902510256
In Proceedings of the 3rd International Conference on Agents and Artificial Intelligence (ICAART-2011), pages 251-256
ISBN: 978-989-8425-41-6
Copyright
c
 2011 SCITEPRESS (Science and Technology Publications, Lda.)

dealing with the large amounts of digital information
available on the Internet.
The EVA architecture has been designed using
an approach that has given some successes for the
study of complex systems (Heudin, 2007). The first
phase of this approach is a top-down analysis that
defines complexity levels and their related
components. The second phase is a bottom-up multi-
agent simulation that attempts to capture the
behavioural essence of the complex phenomena. The
idea is that the complex properties that cannot be
simulated using a classical approach will be likely to
emerge from the interactions between the agents. If
defined and organized correctly, the resulting system
should exhibit the appropriate dynamical
behaviours. In the case of EVA, the emerging
property we are looking for is machine intelligence.
2.2 Nano-agent Architecture
A typical EVA application is composed of one or
more “nano-agents”, and possibly up to a large
number if necessary as in natural swarms (Heudin,
2009). We call them “nano” because of their small
size and resource requirement compared to most
existing software environments. An application can
be composed of several “execution environments”
running on a network of computers. Each of these
environments includes a set of nano-agents and a
nano-server which diffuses messages locally. In the
current implementation, the core technology is
available both in Java and C. A typical EVA
application must have the following properties
(Langton, 1989):
1. The application is modelled as a dynamical
network of agents.
2. Each agent details the way in which it reacts to
local situation and interactions with other agents.
3. There is no agent that directs all the other agents.
4. Any behaviour or global pattern is therefore
emergent.
Such a multi-agent system must also take
advantage of a distributed environment, exploiting
hierarchy and concurrency to perform large-scale
computation.
2.3 NanoScheme Language
EVA provides researchers and developers a user
friendly language which is embedded in each nano-
agent. This language, called nanoScheme is based on
Scheme (Sussman, 1975) and inspired by the RISC
approach (Heudin, 1992). It includes a reduced set
of primitives which is a subset of the Scheme R4RS
specification (Clinger & Rees 1991). Most of the
missing features of the Scheme specification could
be added by programming them directly in
nanoScheme.
2.4 Bio-inspired Core
The nanoScheme language includes also a reduced
set of bio-inspired primitives. They have been
designed in the same spirit of Tom Ray’s Tierran
assembly language (Ray, 1991). That is, the
production of synthetic organisms based on a
computer metaphor of organic life in which CPU
time is the “energy'” resource and memory is the
“material” resource. For example, the reproduce
primitive creates a new nano-agent in the local
environment, and the diffuse primitive diffuses a
message to all nano-agents in the local environment.
The complete set of bio-inspired primitives has been
described in (Heudin, 2010).
Note that the remote execution of code on distant
nano-agents is a natural feature of the language by
simply diffusing nanoScheme expressions. These
expressions are then evaluated by all nano-agents.
This approach enables an easy implementation of
distributed algorithms on nano-agents.
3 SCHIZOPHRENIC APPROACH
3.1 Believable Character
Traditionally, virtual characters were mainly
designed using a computer graphics approach in
which visual realism is the ultimate goal. Therefore
most researchers have looked at believability from
the visual perspective (Vala, 2002). On the other
hand, research in conversational systems has gained
much attention in recent years (Allen, 2001), but
most underlying characters are too simple to produce
a believable conversation. In other words, there is a
lack of personality, back-story and emotions during
dialogues. Real-life characters are far more complex
and versatile than the straightforward character
model of most conversational agents.
In order to create more realistic characters, we
can learn from scenarists and novel writers since
believable characters are the essence of successful
fiction writing (McCutcheon, 1996). Writers are
always looking for new ways to create believable
characters, but most of them use models based on
some stereotypes, personality types or traits. For
example, each type can have its own heroic traits
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
252

and fatal flaws that will bring these characters into
conflict with other people and with themselves. One
of the well-known model of personality traits, used
both in psychology and novel writing, is the “Big
Five” factors (or Five Factor Model) which enable to
describe human personality using five broad axes:
openness, conscientiousness, extraversion,
agreeableness and neuroticism (Digman, 1990).
Each factor consists of a cluster of more specific
traits that correlate together. For example,
extraversion includes such related qualities as
sociability, excitement seeking, impulsiveness, and
positive emotions. The use of this kind of model
enables to create characters with more depth and
dimension.
3.2 Multiple Personalities
There are many models of personality traits, each
one with their own advantages and applications.
Rather than choosing a specific model and thus a
single fixed profile, we prefer to construct a
character identity as the emerging property of
several arbitrary personality traits. The idea is that
the coherent behaviour (or sometimes not so
coherent) of a person results from the many facets of
its personality which are themselves the results of its
own history. This gives the character designer the
ability to compose rich and versatile personalities
without constraints in terms of number or type of
traits.
We have called this approach “schizophrenic”
because the character’s identity is composed of a set
of distinct personalities, each with its own pattern of
perceiving and interacting with the user (Heudin,
2009). Note that this term is used here as a metaphor
since the accurate psychological term for mental
illness with multiple personalities is Dissociative
Identity Disorder, not schizophrenia.
3.3 Emerging Intelligence
In the schizophrenic approach, each individual
personality that composes the character’s identity is
implemented as an autonomous EVA nano-agent.
All nano-agents receive messages from the user and
the other nano-agents, thanks to the diffuse function.
As introduced in section 2.4, this primitive roughly
works like the diffusion of chemical information
between biological cells. Thus, all nano-agents are
able to react to the user’s input by computing an
appropriate answer message. These messages are in
turn diffused to the other personality traits. As a
rough but convenient metaphor with the brain, we
have called these messages “thoughts”.
After this first phase of nano-agents diffusing
thoughts, the resulting complex system of interacting
personalities must converge in order to produce a
coherent answer. There are many possible
approaches for “reconnecting” personalities of the
disparate alters into a single identity. One could be
to let the system self-stabilize toward an attractor.
Another way is to select after some time the thought
with the highest evaluation using a score-based or
fitness-based approach.
4 MS. HOUSE EXPERIMENT
4.1 Character Design
In this section, we present an experiment called Ms
House based on the previously described
schizophrenic approach. Ms. House is a virtual
psychoanalyst specialized in the evaluation of
intelligence. The user can interact with Ms House in
natural language in order to have an evaluation of
his IQ score (Intelligence Quotient). However, Ms
House must be considered more like a casual game
than a real and accurate IQ test. The character design
of Ms House is inspired by two pre-existing
characters.
The first one is the famous Dr. House from the
eponym TV series created by David Shore (Shore,
2004). Dr. Gregory House is an unconventional
medical genius who heads a team of diagnosticians
at the fictional Princeton-Plainsboro Teaching
Hospital.
The second character is Minna Bernay who was
the younger sister of Martha Bernays, Sigmund
Freud’s wife. According to descriptions of her, she
was an intelligent woman, with a lively personality
and a sense of humour. On occasion she could be
highly sarcastic and caustic at times. Sigmund
Freud’s relationship to Minna Bernay has given rise
to considerable speculation (Hirschmüller, 2005).
Ms. Minna House’s character is thus a genius
psychoanalyst with specializations in evaluation of
intelligence who uses unconventional methods. She
is brilliant and she impresses her patients with
accurate but cynical diagnosis and bitter sentences.
Therefore, Ms House is a rich and typed character
compared to most existing conversational agents.
4.2 Ms House’s Inner Personalities
In order to create such a rich and complex character,
Ms House is composed of 12 autonomous nano-
A SCHIZOPHRENIC APPROACH FOR INTELLIGENT CONVERSATIONAL AGENTS
253

agents:
1. Minna: this personality implements Minna
House’s main character profile and back-story.
2. House: this personality implements Dr. House’s
famous way of speaking using an adaptation of the
TV Series screenplay and dialogues.
3. Eliza: this personality is an implementation of
the Eliza psychiatrist program developed by Joseph
Weizenbaum.
4. Artificial: this personality “knows” that Ms
House’s is a conversational agent. It reacts like an
artificial creature.
5. Intelligence: this personality includes a
knowledge base for answering questions about
intelligence and artificial intelligence.
6. IQ-test: this personality evaluates the
intelligence of the user by asking typical IQ test
questions.
7. Profiler: this personality evaluates the user’s
personality using the “Big Five” model by asking
specific questions.
8. Neutral: this personality implements a neutral
and calm character with common language answers.
9. Default: this personality makes a default answer
to any question or sentence.
10. Silent: this personality reacts when the user
waits too much time or does not answer.
11. Logger: this nano-agent logs the conversation
between Ms House and the user.
12. Schizophrenic: this nano-agent selects a thought
and sends it to the user’s interface.
4.3 Implementation
Each personality is implemented as an independent
nano-agent using the nanoScheme language and
dedicated natural language processing features such
as categories extraction and template expressions
(Heudin, 2010). These functions allow the design of
efficient behavior rules for implementing natural
language interactions with the user. The following
code gives a simple illustrating example of the use
of these functions:
; create a list of keywords associated with
the BYE category
(category “Generic” “BYE” '(
“bye” “goodbye” “see you” “ciao”))
; create a list of template answers
associated with BYE
(template “BYE” '(
“Bye bye.”
“Goodbye human being.”
“It was a pleasure to discuss with
you.”))
; create a rule handling the way to answer
to most kinds of “bye” sentences
(rule “goodbye” 2
; condition part
'(find? *categories* “BYE”)
; action part
'(begin
(diffuse 0.5 HAPPY)
(random-template “BYE”)))
Nano-agents diffuse messages in the execution
environment using a simple format:
[(message name value body)]
The brackets define the beginning and the end of the
message. Each message is a nanoScheme expression
where name is the identifier of the sender, value is a
real number and body is the content of the message.
For example, the following an example of a message
sent by the House nano-agent:
(message “House” 100.5 "Sorry, it's not nice
to laugh at mad people.")
In this experiment, the schizophrenic nano-agent of
Ms House implements a simple scheme for choosing
the answer among the thoughts sent by the different
personalities. It is an adaptation of the Fitness
Proportionate Selection algorithm used in Genetic
Algorithms and Evolutionary Computation, also
known as Roulette-Wheel Selection (Baker, 1987).
This could be described similar to a roulette wheel in
a casino where a proportion of the wheel is assigned
to each of the possible selections based on their
fitness value. In Ms House, the fitness value is
replaced by a weight value attached to each
message. Then a random selection is made similar to
how the roulette wheel is rotated. While candidate
thoughts with a lower weight will be less likely to be
selected, there is still a chance that they may be.
4.4 Ms House, Eliza and Alice
In this section, we report a part of a typical
interaction experiment with Ms House, an
implementation of Eliza (Weizenbaum, 1966) and
the Alice chatterbot (Wallace, 2002). This
experiment is based on a predefined dialogue
scenario. For each user input, we show below the
answers from Eliza, Alice and Ms House:
(1) User: Hi.
Eliza: Hi. What seems to be your problem?
Alice: Hello there.
Ms House: Okay. Lie back and lift up your sweater
please.
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
254

(2) User: How are you today?
Eliza: Fine. Thanks
Alice: My logic and cognitive functions are normal.
Ms House: I'm the doctor. You're the patient. Don't
bother me.
(3) User: What is your name?
Eliza: Why are you concerned over my name?
Alice: My name is ALICE.
Ms House: My name is Doctor Minna House.
(4) User: How old are you?
Eliza: What do you really want to know?
Alice: I am the product of over five years' research.
Ms House: Just an advice: never ask such a question
to a woman.
(5) User: Are you some sort of stupid program like
Word?
Eliza: Why are you so interested whether I am some
sort of stupid program like Word or not?
Alice: I don't know whether or not I am some sort of
stupid program like Word. I am a chatterbot
developed by Dr. Richard S. Wallace.
Ms House: Is it still illegal to perform an autopsy on
a living person?
(6) User: Sorry.
Eliza: What does that suggest to you?
Alice: There's no need to apologize.
Ms House: I prefer that.
(7) User: I’m so sorry.
Eliza: Perhaps you would like to talk about a
problem?
Alice: Do you really mean it?
Ms House: Alright. Sometimes I have unexpected
runtime errors too.
(8) User: Can you help me?
Eliza: Why do you think I couldn't help you?
Alice: What kind of help would you like?
Ms House: You believe I can help you. Don't you?
(9) User: What is your problem?
Eliza: So why are you talking to me?
Alice: I don't have a problem.
Ms House: Sorry, it's not nice to laugh at mad
people.
(10) User: What is my problem?
Eliza: All people say this to me in the beginning.
Alice: A very serious problem?
Ms House: Do you usually place yourself nearer to
the side than in the center of the room?
(11) User: Sure.
Eliza: I understand.
Alice: OK let's do it. Oh.
Ms House: May be the best way is to have an idea of
your IQ. I'll ask some questions during our discus-
sion. Okay?
[…]
(12) User: Thank you.
Eliza: Thank you.
Alice: You are quite welcome.
Ms House: You’re welcome.
(13) User: Bye.
Eliza: Goodbye.
Alice: Sayonara.
Ms House: Well... I finally got the last word.
5 DISCUSSION
We can note the quality of answers of the Alice
chatterbot compared to those of Eliza. The only
problem with Alice seems the answer for question
(11). Eliza is generally rephrasing the user’s entries,
as for questions (3) and (5), or selecting a random
default sentence as for questions (7), (9) and (11).
The answers made by Ms House reflect its
humoristic and cynical character while keeping a
good balance with its other personality traits:
answers (1), (2), (5), (9) and (13) were given by the
House nano-agent, answers (3) and (4) were given
by the Minna nano-agent, answers (6) and (12) were
given by the Neutral ano-agent, answer (7) was
given by the Artificial nano-agent, answer (8) was
from the Eliza nano-agent, answer (10) was from the
Profiler nano-agent and (11) from the IQ-test nano-
agent. After question (11), Ms House took then the
initiative of the conversation by asking typical IQ
test questions to the user. We have not reported this
phase here (noted […]), since it was not relevant to
compare Ms House with Eliza and Alice which are
only answering questions with no underlying
motivation or goal.
Even if Ms House’s qualitative results seem very
good in this example, this is not always the case. In
some cases, the emerging behaviour is not so
coherent. This is due to the stochastic selection
scheme which is based on fixed weights attached to
each personality traits. We think that this problem
could be solved by using dynamical weights more
related to the context of the conversation. This could
be achieved by an emotional metabolism (Gebhard,
2005) (Heudin, 2010) and a contextual short-term
memory that could alter the initial weight balance
between personality traits.
The second problem concerns the openness of
the knowledge base and learning capabilities as
mentioned in the introduction. We have reported in a
previous study the use of web mining agents in order
to fetch information and use it in the flow of
conversation (Millet, 2007). In addition to this, we
think that it will be interesting to experiment the
A SCHIZOPHRENIC APPROACH FOR INTELLIGENT CONVERSATIONAL AGENTS
255

same approach to the conversation logs. This
growing database could serve as the basis for
learning using the Genetic Programming features of
the EVA architecture (Heudin, 2010). We argue that,
even in the case of no interaction with a user, some
dedicated nano-agents could learn new information
and behaviours from both the web and the log
database.
6 CONCLUSIONS
We have presented in this paper a novel approach
for creating intelligent conversational agents. This
approach relies both on a more sophisticated
character design inspired by novel writer practices
and its implementation using a “schizophrenic”
model. The latter represents a promising scheme
thanks to its intrinsically parallel and bio-inspired
features. On a long term, we think that we can create
an intelligent conversational agent based on a swarm
rather than a small number of personalities.
While our theoretical framework is based on the
complex system approach, our experimental
approach focuses on real-world applications. Our
approach has obvious applications for designing
intelligent agents for commercial web sites and
marketing studies. We also like to imagine virtual
assistants on smart phones, assistants for lone aged
and/or sick people, for learning foreign languages,
virtual characters in video games, for robotic and
embedded applications.
REFERENCES
Allen, J., Byron D., Dzikovska, M., Ferguson, G.,
Galescu, L., Stent, A. 2001. Towards conversational
human-computer interaction. AI Magazine, vol. 22.
Baker, J. E., 1987. Reducing Bias and Inefficiency in the
Selection Algorithm. In Grefenstette, J. J. (ed.):
Proceedings of the Second International Conference
on Genetic Algorithms and their Application,
Lawrence Erlbaum Associates, Hillsdale, pp. 14-21.
Cassel, J., Sullivan, J., Prevost, S., Churchill, E. F., (eds.)
2000. Embodied Conversational Agents. MIT Press,
Boston.
Clinger, W., Rees, J. (Eds.) 1991. Revised(4) Report on
the Algorithmic Language Scheme.
Digman, J. M., 1990. Personality structure: Emergence of
the five-factor model. Annual Review of Psychology,
41, pp. 417-440.
Gebhard, P., 2005. ALMA: A layered model of affect. 4th
International Joint Conference on Autonomous Agents
and Multi-agent Systems, pp. 29-36.
Heudin, J.-C., Panetto, C., 1992. RISC Architectures,
Chapman & Hall, London.
Heudin, J.-C., 2004. Evolutionary Virtual Agent.
Proceedings of the Intelligent Agent Technology
Conference, ACM-WIC-IEEE, Beijing, pp. 93-98.
Heudin, J.-C., 2007. Modeling Complexity using
Hierarchical Multi-Agent Systems. Proceedings of the
6th International Workshop on Data Analysis in
Astronomy, Erice.
Heudin, J.-C., 2009. An Evolutionary Nano-Agent Control
Architecture for Intelligent Artificial Creatures.
Proceedings of the 14th International Symposium on
Artificial Life and Robotics, AROB, Beppu.
Heudin, J.-C., 2010. A Bio-inspired Nano-Agent
Architecture for Intelligent Agents. In Web
Intelligence and Intelligent Agents, Z. Usmani (Ed.),
In-Tech, Olajnica, pp. 1-11.
HirschMüler, A., 2005. http://www.answer.com/topic/
bernays-freud-minna
Langton, C. G., 1989. Artificial Life, Proceedings of the
First Artificial Life Conference, Vol. VI., Addison-
Wesley, Redwood City.
McCutcheon, M., 1996. The Writer's Digest Sourcebook
for Building Believable Characters, Writers Digest
Books, 1st edition.
Millet, P., Heudin, J.-C., 2007. Web mining in the EVA
intelligent agent architecture, Proceedings of the
Intelligent Agent Technology Conference, ACM-WIC-
IEEE, California.
Ray, T. S., 1991, An approach to the synthesis of life.
Proceedings of the 2nd Artificial Life Conference,
Addison-Wesley, Redwood City, pp. 371-408.
Shore, D., 2004. http://en.wikipedia.org/wiki/House.
Sussman, J. G., Steele, G. L., 1975. Scheme: An
Interpreter for Extended Lambda Calculus, MIT AI
Lab Memo AIM-349.
Vala, M., Paiva, A., Gomes, M. R., 2002. From Virtual
Bodies to Believable Characters. AISB Journal, 1(2).
Wallace, R.S., 2002.
The anatomy of Alice,
http://www.alicebot.org.
Weizenbaum, J., 1966. Eliza, A Computer Program for the
Study of Natural Language. Communication between
Man and Machine, Communication ACM, 9, pp. 36-
45.
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
256