Saturday, April 30, 2011

The Media Equation and Artificial Creatures

I want to make a short comment about “The Media Equation” from Reeves and Nass* since it has many implications for the design of artificial creatures.
Their theory claims that humans are social animals that come from their evolution history: they interact following the rules of interpersonal communication even if the interaction is not with a person (anthropomorphism). In other words, they tend to respond to media as they would another person like by being polite, cooperative, and attributing a personality type.
*Byron Reeves & Clifford Nass, The Media Equation: How People Treat Computers, Television, and New Media like Real People and Places, Cambridge University Press (1996).

More precisely, the media equation relies on eight main propositions:
(1) Everyone responds socially and naturally to media.
(2) Media are more similar than different.
(3) Reactions occur automatically without conscious effort.
(4) When using any type of media, a person is likely to assign it a personality.
(5) What seems true is more important than what is true.
(6) People respond to what is present.
(7) People like simplicity.
(8) People already know how to function in the natural world (be polite, etc.).

From my experience, I think this is also true for artificial creatures. It appears to me that the Media Equation is a good guide for creating interacting characters. It does not contain all the rules, but these ones are important.

Des agents conversationnels à la Société Générale

J'ai participé mercredi à la rencontre organisée par Pacte PME à la Société Générale (La Défense) en tant qu'expert sur le thème de l'IA et du langage naturel. Le thème central était plus précisément celui des agents conversationnels et de leurs applications. Quatre entreprises invitées ont présenté leur offre produits et services à une assemblée composée d'une quarantaine de cadres du groupe : Davi Interactive, Dialonics, Spirops et Zenvia. Je connaissais bien Spirops puisque son créateur, Axel Buendia, a développé la version initiale de sa technologie dans le cadre de sa thèse au laboratoire de l'IIM. Je connaissais moins bien les trois autres acteurs, mais ils ont ensemble démontré la vivacité des PME françaises sur ce sujet. Après les présentations, une table ronde de quatre experts, deux du groupe Société Générale, et deux externes (dont votre serviteur) ont réagis aux présentations des entreprises et répondu aux questions du public. Comme l'a souligné fort justement l'un des participants à cette rencontre, il serait nécessaire que les PME spécialisées sur ce sujet prometteur fassent cause commune plutôt que de luter les unes contres les autres sur un marché encore émergent. Leur taille étant généralement assez réduite, elles ne pourraient que profiter d'une alliance ou d'un regroupement qui rendrait plus crédible leurs offres (souvent complémentaires) auprès des grands comptes.
Le second thème était celui de l'Internet des objets avec d'autres entreprises et experts. Pour ma part, j'aurai bien liés ces deux sujets, car je pense que l'avenir est aux objets "vivants", autonomes et intelligents, capables de dialoguer avec les utilisateurs en langage naturel.

Wednesday, April 20, 2011

Complex Artificial Intelligence at Erice

Following my previous post on an extended version of Uncanny Valley applied to the study of complex systems, I want now to explain with more details my new approach to Artificial Intelligence. I've called this approach "Complex Artificial Intelligence".
There have been numerous models of intelligence. The field of AI has been one of the most prolific since its birth in the middle 1950s. We can roughly class them in four categories:
Symbolic AI: In the early years of AI, most models were based on the theory that human intelligence could be reduced to symbol manipulation. The main outcome of this approach was the development of expert systems based on rule inference and symbolic processing.
Neuron-based AI: The study of artificial neural networks began in the decade before the field AI research was founded. It was based on a formal model of the neuron cell in the form of a simple threshold automaton. Researchers focused on elaborating different types of neural networks and their related learning algorithms. Pattern recognition is the main application field for this approach.
Reactive AI: In the 1990s, behavioral robotic architectures and Artificial Life focused on “intelligence without representation” instead of the classical symbolic computational models. The idea was to work on bio-inspired adaptive systems addressing their basic perceptual and sensorimotor tasks.
Distributed AI: During the same period of time, many projects were dedicated to the development of Distributed Artificial Intelligence and Multi-Agent Systems. The key concept is the abstraction called a software agent: a virtual autonomous entity that has an understanding of its environment and acts upon it. An agent communicates with other agents to achieve a common goal, that one agent alone could not achieve.
While taking advantages of the previous ones, Complex AI considers machine intelligence as a global property that emerges from the large amount of non-linear interactions between numerous small software agents and the environment at the edge of order and chaos.
In the workshop “Image In Action” at Erice (Sicily), I have illustrated this approach using three models: (1) a generalized Game of Life Cellular Automata showing the emergence of complex behaviors at the edge of chaos; (2) the emergence of simple “intelligent” behaviors in the evolution of species in the Lifedrop virtual world; and finally the “schizophrenic” model of the Ms House experiment using EVA (Evolutionary Virtual Agent). Ms House is an intelligent character composed of multiple distinct personalities. Each individual personality is implemented as an autonomous bio-inspired nano-agent which is able to react to the user’s input by computing an appropriate message. A coherent identity emerges from the interactions between these nano-agents and the user at the edge of order and chaos. Ms House is composed of near 30 different personalities at this time. Our goal is to have a swarm of over 100 personalities in the near future.

Friday, April 08, 2011

The uncanny valley and complex systems

In two weeks I will make akeynote speech at the "Science: image in action", the 7th International workshop data analysis in astronomy at the Ettore Majorana Foundation and Centre for Scientific Culture in Erice, Italy.
The title of my talk is "Reality, models and Representations: the case of Galaxies, Intelligence and Avatars." I will present three different case studies: the modeling of colliding galaxies in astronomical research, the modeling and simulation of human intelligence in computer science, the creation of digital avatars in computer graphics. I will clarify the relationships between reality, models and representations using a semiotic approach based on the classical triangle of references. I will show that visual representation is as important as the modeling and that there is no possible equivalence between the real phenomena and its model and representation for complex systems.
The main reason is a “complexity barrier” between the real phenomenon and its synthetic counterpart. This barrier is due to the huge difference in terms of quantitative and qualitative complexity. In addition, I think that the “uncanny valley” problem, originally formulated by M. Mori (see figure) for humanoid robots, can be generalized to all modeling and simulation of complex systems. The importance of details in visual representations and behaviors increases exponentially when it becomes close to the one of a real phenomena. In other words, if a very small detail can transform an empathic avatar into a monster, it can also substantially decreases the benefits of a model and its representation.