Master 2 Internship and PhD Subjects*

Most studies on planning and reasoning assume that the world is completely and correctly known and that actions are deterministic with no duration and no failure. The aim is to extend existing algorithm in order to handle this framework.

Références:

1. Florence Dupin De Saint Cyr, Jérôme Lang. Belief extrapolation (or how to reason about observations and unpredicted change). In : Artificial Intelligence, Elsevier, Vol. 175, pp. 760-790, January 2011.
2. Florence Dupin De Saint Cyr. Scenario Update Applied to Causal Reasoning . In : International Conference on Principles of Knowledge Representation and Reasoning (KR 2008), Sydney, Australia, 16/09/2008-19/09/2008, Gerhard Brewka, Jerome Lang (Eds.), AAAI Press, pp. 188-197, September 2008.
3. Matthew Molineaux and David Aha. Learning Unknown Event Models. In : AAAI. pp. 395-401, 2014.
4. Matthew Molineaux, Matthew Klenk and David Aha. Goal-Driven Autonomy in a Navy Strategy Simulation. In: AAAI. pp 1548-1554, 2010.
5. Dustin Dannenhauer, Hector Munoz-Avila and Michael Cox. Informed Expectations to Guide GDA Agents in Partially Observable Environments. In : IJCAI, pp 2493-2499, 2016.

More precisely, our cognitive model of agent acceptance is described on the Figure bellow:

On the right hand side of this Figure we consider a human agent having its logical knowledge base (with factual observations about the world and generic statements such as Miradoux is a wheat variety, wheat contains proteins), its own associations (proteins are related to nutrition) and its opinions (I like Miradoux, I don’t like spoiled wheat). The logical knowledge is stored as a logical knowledge base (I, in the Figure) expressed in Datalog+/- (for re-utilisability purpose).

The associations (II) (that can relate two pieces of information as well as a piece of information to an appreciation—in that case it is called opinion) are more difficult to elicit and are not often handled in the literature. One of the main difficulties is that the associations depend on the profile of the person (expert, non-expert, etc.).

A third parameter is also required by our model, namely, the cognitive availability (abbreviated ca) of the agent (III) which depends on the agent’s interest about the particular argument and on the amount of attention he has to spent (its precise definition is not studied here, it may be based on the agent mood, her knowledge, sometimes on the speaker, on the topic of the argument, etc.). This cognitive availability is a parameter that we use to filter the possible reasoning the agent is able to do.

On the left hand side of the figure, we show the proposed process of argument acceptance using the cognitive model. When the agent hears a new argument [step (1)], a number of critical questions are fired [‘‘is the premise of the argument correct?’’, ‘‘do I agree with the conclusion?’’, ‘‘can I infer is the premise of the argument correct?”, “do I agree with the conclusion?”, “can I infer the conclusion from the premise based on what I know?” - step (2)).

Thanks to the proposed cognitive model, it will then be possible to compute reasoning paths (i.e. sequences of logical rules and association rules constituting a chain of inferences that leads to a desired conclusion) for each critical question (step (3)). For a reasoning path we introduce the notion of effort (cognitive effort to use the rules of the reasoning path) which is confronted to the cognitive availability of the agent. An association is usually effortless while logical reasoning is considered as more expensive. The cognitive availability of the agent allows us to have an upper bound of the effort the agent is able to put into her reasoning paths. The reasoning paths will be selected based on the effort needed to carry on (step (4)). Based on this selection we can accept or reject an argument. Note that the reasoning paths are constructed from the logical knowledge base and the associations that computationally represent the knowledge of the expert.

Références:

1. Pierre Bisquert, Madalina Croitoru, Florence Dupin de Saint-Cyr, Abdelraouf Hecham. Formalizing Cognitive Acceptance of Arguments: Durum Wheat Selection Interdisciplinary Study. In : Minds & Machines, Springer, February 201.
2. Pierre Bisquert, Madalina Croitoru, Florence Dupin de Saint-Cyr, Abdelraouf Hecham. Substantive irrationality in cognitive systems (short paper). In : European Conference on Artificial Intelligence (ECAI 2016), La Hague, Hollande, 29/08/16-02/09/16, Maria Fox, Gal Kaminka (Eds.), IOS Press, pp. 1642-1643, 2016.
3. Pierre Bisquert, Madalina Croitoru, Florence Dupin de Saint-Cyr, Abdelraouf Hecham. Towards an Understanding of Human Persuasion and Biases in Argumentation. In : Cardiff Argumentation Forum, Cardiff UK, 06/07/16-07/07/16, Federico Cerutti , Francesca Toni (Eds.).

Nowadays the transmission of ideas mainly goes through writing. The written language must meet some syntactic and grammatical rules, with a reading direction of the words which are the building blocks of sentences that should be read in a sequential order. Even if it allows us to express complex and subtle ideas, some representations go beyond the linear mode by adding a spatial dimension to information, possibly on an interactive support, or an approach combining a comprehensive view and a detailed one.

They already exist a lot of visual representation frameworks that are more or less commonly used : Mind maps introduced by Tony Buzan, Concept maps, Venn diagrams, Historical timelines, Programming flowcharts, Geographical maps... However they all have some drawbacks.

In order to overcome these drawbacks, we have formalized four principles and expressed eight postulates as a first step towards the definition of a theoretical system that could help to characterize user-friendly and efficient visual representation languages. Moreover, we have proposed a new language called VTL, for visual typed representation language, which can satisfy these postulates.

The next directions of research are the development of a graphical user interface (GUI) specific for VTL. The study of the automatic translation of VTL into a formal non visual language in order to propose inferences and consistency checks. Another direction of work would be to study how VTL allows to encompass links that were not handle by MOT, namely RCC8 relations or Allen intervals, or other relations between concepts (we could refer for instance to the linking-words typology written by Christian Barette).

Références:

1. Florence Dupin de Saint-Cyr, Denis Parade. Représentation des Connaissances dans un Langage Visuel Typé: une première approche. Journées Francophoes d'Ingéniérie des Connaissances (IC 2018), Nancy, AFIA, pp 179-194, 2018
2. Florence Dupin de Saint-Cyr, Denis Parade. Knowledge Representation in a Visual Typed Language: from Principles to Practice. International Conference on Information Visualisation (IV 2018), IEEE Computer Society, page 350-355, 2018<\li>

We have proposed a decision support framework (called BLF) that is visually described as a bipolar graph of decision principles where decisions are associated with utilities. This framework makes it possible to express both the characteristics of a decision problem and reasoning in order to select the appropriate decision corresponding to the situation.

The objective of this internship is to create automatically the BLF from databases. The techniques used will be learning techniques such as learning of rules, and their exceptions (called inhibitors), rule clustering. We would like to explore the links with formal concept analysis.

References:

1. Florence Dupin de Saint-Cyr, Romain Guillaume, Umer Mushtaq Explainable Decisions under incomplete knowledge with supports and weights. IEE International Conference on Fuzzy Systems (FUZZ-IEEE 2019), pp 1-6, 2019
2. Florence Dupin de Saint-Cyr, Romain Guillaume. How Pontential BLFs Can Help to Decide under Incomplete Kmowledge. International Conference on Information Processing and Management of Uncertainty in Knowledge-based Systems (IPMU 2018), pages 86-98, 2018