U bent hier


Opgelet! Dit event heeft al plaatsgehad.

Multi-Agent Causal Models: Learning and Inference

maandag, 19 december, 2005 - 17:00
Campus: Brussels Humanities, Sciences & Engineering campus
Faculteit: Science and Bio-engineering Sciences
Sam Maes

In this dissertation we have introduced multi-agent causal models (MACM).
These are a graphical modeling technique where several agents model overlapping parts of
the domain. Each agent models his own subdomain with a semi-Markovian causal model,
i.e. a causal model that allows a certain type of unobservable variables.

Each of the individual agent models in a MACM has private variables that it keeps
confidential, and public variables that it shares in an intersection with another agent.

Multi-agent causal models extend existing techniques in the following two ways.

First of all, single agent causal models, such as causal Bayesian networks, are extended to
the multi-agent case.
The motivation for doing this is that domains to be modeled become evermore complex.
Therefore, models that represent such systems with a single centralized approach become
hard to learn and maintain.
Instead, several small models that represent a modular part of the domain are a more
tractable way to represent complex systems.
Furthermore, components of complex systems, such as a large industrial plant, can be
spatially remote from each other. Modeling such a system with a single agent approach
makes the model vulnerable to communication problems.
Situating an agent near each remote component of a system can alleviate some of these
problems, as it can first process local information and then only communicate the most
relevant information to other agents. Furthermore, if communication completely fails, the
individual agents can assure some minimal functionality in anticipation that the
communication is restored.
Finally, a multi-agent approach can exploit the fact that some calculations in the model
can happen in parallel.

Secondly, multi-agent probabilistic models, such as multiply sectioned Bayesian networks,
are extended to allow for causal inference.
The motivation for that step is that causal inference allows to predict the effect of changing
the domain that is being modeled.
This is especially useful as a component in decision support systems, which are computerbased
systems that help human users to make complex decisions. Causal inference
techniques can help a user to assess the effect of his possible actions, before actually
performing them.

Next to introducing multi-agent causal models, we have also developed algorithms for
performing multi-agent causal inference and learning the structure of these models.