Interaction intelligente par coordination sans communication préalable entre des agents logiciels et des êtres humains dans le cadre de tâches complexes impliquant une criticité des situations.
Martial Razakatiana · HAL (Le Centre pour la Communication Scientifique Directe) · 2021
This thesis contributes to the modeling of intelligent human-computer interaction, by considering a multi-agent assistance system in the context of the cooperative realization of one or more complex tasks.We address a problem concerning a coordination mechanism in a human-machine system as an interaction model between one or more human operators and one or more assistive software agents. In particular, we consider an interactive system in which a complex task involving a criticality of the situation requires a cooperation between human(s) and software agent(s). The proposed interaction model is based on game theory. Thus, each participant is modeled as a player who plays at each iteration in a dynamic way. We have characterized the human beings by their experience level and their workload level at each iteration of the game. For software agents, some expertise rules are exploited. They can make decisions according to the situation and the values ofthe variables involved: to cooperate or not to cooperate. For human beings, they can make any decision at a given time according to their experience and skills. The software agents must help the humans through cooperation or let them work autonomously, in favor of availability for another task. Thanks to the game payoff matrix and Nash equilibrium, each of the software agents can intervene or take care of another task, thus optimizing their interventions. This approach allows us to consider a context-sensitive system, either at the level of the user or the contexts of the interaction environment.For the case study, we developed a road traffic simulator based on a traffic model available in Netlogo, a modeling environment for multi-agent system development, using the principle of participatory simulation. The simulator allows the interaction between human beings and software agents in the context of traffic control through traffic lights. At first, we considered a single human and a single assistant agent. Then, two humans and one assistant agent were considered. Finally, a study involved two humans and two assistant agents. The analysis of the results after the execution of the simulator for the three case studies shows the interest and the feasibility of our model. These first promising results also allow us to study the efficiency of agents and humans in the collaborative realization of a complex task. Several avenues for future research are now considered.