ABSTRACT. I will present several of our works regarding modeling social and emotional behaviors for embodied conversational agents. Lately we have used methods that rely on corpus analysis. We have gathered a corpus of more than 80 dyads. This corpus is annotated on different levels, high level such as impression of competence or social attitude, and multimodal behaviors. By applying sequences mining we extract behavior patterns involved in the change of perception of an attitude or competences. In this talk I will also present our work on simulating group of agents interacting with each other. In particular I will describe the turn-taking management mechanisms we have developed that take into account social attitude agents have toward each other.

ABSTRACT. Problem simplification is a topic of continuing interest in the field of constraint satisfaction. In this paper we examine properties associated with the basic idea of substitutability and show how certain forms of substitutability can be organized into a strict hierarchy. One of these properties, here called replaceability, has been identified by other authors as being of special interest. In this work we confirm these earlier claims and show that replaceability is significant because it is the most general property in the hierarchy that allows inferences from local to global versions of this property. To make use of this discovery, we introduce two algorithms for establishing neighbourhood replaceability, and we present an initial experimental examination of these algorithms including ways to improve their performance through ordering heuristics and various kinds of inference.

ABSTRACT. Two non deterministic algorithms for generalizing a solution of a constraint expressed in second order typed $\lambda$-calculus without constants are presented. One algorithm derives from the higher order unification rules by D. C. Jensen and T. Pietrzykowski, the other is abstracted from an algorithm by N. Peltier and the author for generalizing proofs. A framework is developped in which such constrained generalization algorithms can be designed, allowing a uniform presentation for the two algorithms. Their relative strength at generalization is then analyzed through some properties of interest: their behaviour on valid and first order constraints, or whether they may be iterated or composed.

ABSTRACT. We apply genetic algorithms (GAs) to evolve cyclic finite automata for scheduling the dispatch of trucks in mines. The GA performs well generally, and on problems which include one-lane roads, the GA was able to find solutions that utilised shovels very well, with low contention and using fewer trucks than both the widely-used linear programming DISPATCH algorithm, and commonly-used greedy heuristics. The GA provides significant cost-savings, or production increases, on problems where alternative algorithms do not adapt well.

ABSTRACT. Nature showcases swarms of animals and insects performing various complex tasks efficiently where capabilities of individuals alone in the swarm are often quite limited. Swarm intelligence is observed when agents in the swarm follow simple rules which enable the swarm to perform certain complex tasks. This decentralized approach of nature has inspired the artificial intelligence community to apply this approach to engineered systems. Such systems are said to have no single point of failure and thus tend be more resilient. The aim of this paper is to put this notion of resilience to the test and quantify the robustness of two swarm algorithms, namely ¸"swarmtaxis" and "FSTaxis". The first simulation results of the effects of introducing an impairment in agent-to-agent interactions in these two swarm algorithms are presented in this paper. While the FSTaxis algorithm shows a much higher resilience to agent-to-agent communication failure, both the FSTaxis and swarmtaxis algorithms are found to have a non-zero tolerance towards such failures.

ABSTRACT. Robot navigation in domestic environments still lacks the robustness and legibility that users expect of autonomous robots. This paper shows how a shift from optimizing the two typical modules in navigation (path planning and control) towards a dynamic interaction of more, not necessarily hierarchically linked, modules leads to more robust navigation behavior. This claim is supported by experiments in a simulated household with a simulated PR2 robot.