Search for an Appropriate Behavior within the Emotional Regulation in Virtual Creatures Using a Learning Classifier System

Abstract

Emotion regulation is a process by which human beings control emotional behaviors. From neuroscientific evidence, this mechanism is the product of conscious or unconscious processes. In particular, the mechanism generated by a conscious process needs a priori components to be computed. The behaviors generated by previous experiences are among these components. These behaviors need to be adapted to fulfill the objectives in a specific situation. The problem we address is how to endow virtual creatures with emotion regulation in order to compute an appropriate behavior in a specific emotional situation. This problem is clearly important and we have not identified ways to solve this problem in the current literature. In our proposal, we show a way to generate the appropriate behavior in an emotional situation using a learning classifier system (LCS). We illustrate the function of our proposal in unknown and known situations by means of two case studies. Our results demonstrate that it is possible to converge to the appropriate behavior even in the first case; that is, when the system does not have previous experiences and in situations where some previous information is available our proposal proves to be a very powerful tool.

Figure 1

Environment. Environment example for virtual creatures, showing a possible environment for virtual creatures where O₁,…, O_n are possible objects perceived, V₁,…, V_n are other virtual creatures perceived, and M₁,…, M_n are possible memories recovered from previous situations.

Figure 2

Emotion Regulation Model. A proposed model based on neural evidence and psychological theories (SI = sensory input, AS = affective signal, ES = emotional signal, PA = possible action, EMS = emotional mental state, AB = appropriate behavior, SS = stimulus signal, SEV = stored emotional value, NSEV = new stored emotional value, NM = new meaning, and RB = regulated behavior).

Figure 3

An example of internal emotional action from a virtual creature. Showing possible emotional behaviors from the environment where B₁,…, B_n are possible emotion behaviors and R₁,…, R_n are the possible responses from the environment.

Figure 4

Selection of the rule that best satisfies the conditions of the environment. Showing possible emotional behaviors in response to the environment where B_n are possible emotional behaviors and R_n are the possible responses from the environment.

Figure 5

General process of GXCS. Showing a general process of GXCS with inputs and outputs.

Figure 6

A sample of a rule's structure for LCS. Showing a syntax rule where E1, E2,…, E6 are emotions and ST is the current situation.

Figure 7

A sample of a rule's database for GXCS. Showing a syntax rule where E1, E2,…, E6 are emotions and ST is the current situation.

Figure 8

A sample of generated rules for GXCS. Showing rules generated by an action with 6 components.

Figure 9

Graph of search behavior showing results from the first experiment in three columns. Each column is one emotion; the numbers are the rules generated in the experiment. The time taken for each experiment was less than one second.

Figure 10

Graph of association of behavior. Showing a result of the second experiment, with two columns—unknown situations and known situations—for each experiment using the same LCS. The difference between the unknown and known situation was less than 15 new rules.

Figure 11

Alfred sample. This picture shows how the model of emotional regulation works. Part (a) shows the environment with difference evaluations stored in the memory; part (b) shows Alfred with an expression of fear, which is the first emotion. Applying the LCS to obtain the appropriate behavior, we have a visible decrease of emotion seen in Alfred in part (c). This result is obtained from the appropriate behavior given by LCS and multiple evaluations of emotional objects in the setting using an emotion regulation architecture.