Esprit Project 25197



















i3 Spring Days Abstract




(Workshop on Behavior Planning for Life-Like Characters and Avatars)

Due out January 10, 1999



Doc. Id: CRR External Relations

Partner: CRR




Author: Riccardo Antonini, Amy Brown

Date: 15/12/98

Revised by: Riccardo Antonini

Availability: Restricted

Status: Draft – Version 2.0 – Proposed





Proposal for Participation in the Workshop on Behavior Planning for Life-Like Characters and Avatars:




Riccardo Antonini

Consorzio Roma Ricerche

Via O. Raimondo, 8

00173 Roma (Italy)

tel. 0039-06-7234606





The idea of having agents showing a "believable character" has been introduced by Bates in [Bates 1994]. The possibility of using an evolutionary procedure in order to cast features of a character has been first introduced by the now classical work on "Biomorphs" by Richard Dawkins in [Dawkins 1986]. The work of Dawkins was meant, among other things, to show the power of cumulative selection in evolution. The original work was focused on the evolution of morphological features. In the Amusement project we have successfully implemented a procedure that allows users to cast their avatar’s gestures by means of an "evolutionary procedure". The selective pressure is their personal preference as in Dawkins’ "Biomorphs". This type of systems originally devised by Dawkins [Dawkins 1986], are now called "collaborative systems" in the scientific community [Symposium on Creative Evolutionary System, Edinbourgh, UK, April 6-9, 1999].

We have now well established technique using collaborative systems in order to develop both morphology and gestures of artificial creatures. Quite obviously one of the most interesting applications of those techniques is the development of believable characters in agents as defined in Bates [Bates 1994]. In this paper we want to introduce the possibility to develop "collaborative systems" that not only shape the morphology of avatars and their relevant gestures, but also the behavior of an avatar or "non participating character" (NPCs, Non Participating Characters, have been illustrated in [Coco 1997]). In the case of a "non participating character" we are dealing with a completely autonomous agent that accepts no inputs other than the ones coming (indirectly) from the environment. In the case of an avatar, of course, there is a human behind that leads it. Anyhow even in the case of an avatar a certain degree of autonomy is highly desirable. Because otherwise the players would have to give to the avatar all the information needed for its behavior. In the case of an avatar, for instance, a very simple example is given by the procedure normally used in order to implement a believable walking "gesture". When an avatar is commanded, by means of a pointing device, to displace itself, the user needs not to tell the avatar all the mechanics of walking, because there is an internal program that automatically triggers the gesture "walking". That happens each time the pointing device tells the avatar to displace itself. The gesture "walking" is hence predefined and triggered upon the event "displace".

By the way the gesture walking, in the Amusement project, is generated by means of a "collaborative system" in which the user controls the evolution of a population of replicas of his/her avatar showing "broader walking patterns", chooses the one that he likes most and triggers the reproduction of replicas showing "walking patterns" "sons" of the chosen "walking pattern".

Still the link between the situation (in the example "displacement") and the gesture (in the example "walking") is fixed.

More formally, let’s consider the situation "s" as a stimulus, and the gesture as a response "r". Then we can define the behavior as the ordered pair (s, r). The behavior can be defined as a function "b" that maps S onto R. Where S is the set of all possible Situations, and R is the set of all possible Responses. This definition is the one used in [Kosko 1992]. In [Kosko 1992] this definition is used in order to illustrate what is called, there, "model-free estimation" of behavior. Model-free estimation is used in Kosko in the context of neural and fuzzy system theory. We will extend here the concept of model free estimation of behavior in the context of "collaborative systems".

All that is necessary in order to estimate a behavior, without the need of using an explicit model, is to substitute the classical learning procedure of a neural network, or the design of a fuzzy system, with the evolution of a function "b" codified by a genome. In the Amusement project we have used a genome that codifies for the gestures and for the association of a given situation to the gesture itself.

Of special interest to those of us in the world of games and entertainment is the fact that showing the user’s real mood can be counter-productive. In order to provide an avatar with the ability to hide emotions whenever it is necessary, we have allowed the user to define his/her real mood along with the mood that he/she wants to simulate. The resulting avatar’s expression and actions will depend on how gifted it is in dissimulating. The association of a real mood and the expressed mood is controlled in the Amusement implementation, by means of a deterministic albeit "fuzzy logic" internal model.

As said before "b" maps S onto R. Now in order to take into account the possibility of dissimulating we will slightly modify the definition of "b" letting "b" be "weighted" by a function of dissimulation "d". In the full paper we will describe into details the whole procedure. Meanwhile the implementation of it can be seen in the Amusement project at