Multi-Agent Methods: An Example of an Architecture and its Application for the Detection, Recognition and Identification of Targets - Information Fusion in Signal and Image Processing - page 209

of co-operation [GAT 96]. The fundamental point that characterizes all of these co-
operations, whether they are confrontational, augmentative, or integrative is the accu-
mulation of information with the objective of making a decision.
10.2.4. Decision and uncertainty management
Decisions have to be made at every step of the vision process. In order to make
these choices, the system needs to constantly estimate the confidence in the current
hypotheses, whether as regards to location, detection, focusing or recognition. Confi-
dence depends on a great number of parameters such as:
- the quality of the selected operators;
- the presence of significant descriptors for objects;
- knowledge about the context, the environment.
It integrates very different elements of information, such as statistical, geometric or
kinematic information, for example. In order to do this, different formalisms have been
suggested, such as probabilistic methods [RIM 93], possibilistic methods [NIF 98],
Dempster-Shafer theory [LEF 96, NIG 00], or also fuzzy methods [MEE 00]. These
formalisms make it possible to approach the problem of the dynamics of beliefs by
using inference mechanisms, in order to draw temporary conclusions [FAB 96]. As
Bremond and Thonnat have aptly pointed out, “the most commonly used methods are
probabilistic .... In the context of scene interpretation ...they provide a strict and
rigorous framework that can be applied to any type of uncertainty” [BRE 96].
Graphs constitute a graphic and efficient way of representing the relations that exist
between the different possible states of the system: causality, dependence or temporal
relations, etc. They also turn out to be well-suited tools for structuring knowledge,
thus making it possible among other things to propagate information and uncertainty
in order to reinforce or decrease hypotheses.
This is why we have chosen a representation in the form of Bayesian networks.
This formalism requires a complete definition of the probabilistic model, but provides
a rigorous framework [FAB 96].
10.2.5. Incrementality and learning
Incrementality presents itself on several levels of the system:
- in the construction of the system itself since the population of agents is adapted
depending on the information available and its relevance;
- in the gathering of information since the information is progressively accumu-
lated in a world model which is made available to the population of agents;