Miklós MOLNÁR

  Assistant Professor                                                                                         Version française

Address :
 
INSA
Département informatique
20, Av. des buttes de Coësmes
F-35043 Rennes Cedex, France
Tel : +33 (0) 2 23 23 86 79
Fax : +33 (0) 2 23 23 84 53
Project ARMOR
Campus de Beaulieu,
F-35042 Rennes Cedex, France
Tel : +33 (0) 2 99 84 74 42
Fax : +33 (0) 2 99 84 71 71

email :          molnar@irisa.fr



Research Interests
 



 Research Activities

Steiner Problem in Graphs and Related Works
From the topology point of view, the lower cost multicast diffusion in a network corresponds to a partial minimum spanning tree (named
also minimal Steiner tree). The construction of such a tree being NP-complete, several heuristic approximation algorithms with polynomial
time have been formulated. It is interesting to find good compromises which can decrease the tree length by using a reasonable
computational time to build and manage the connection. Many of the heuristics use the shortest paths to connect the multicast group
members. The computational time of this kind of construction is weak but the length of the resulting trees is not always acceptable. In our
research, we give some scalable approximate construction algorithm of the optimal tree that can diminish the length of the multicast tree
and still be in polynomial time. This construction makes connections like simple minimum Steiner trees between the already existing
sub-trees of the partial spanning tree. Our first proposition is the modification of the well known Kruskal's heuristic ( 1   and   2 ). The same
idea can be applied on the Takahashi-Matsuyama's heuristic ( 3  ).
Our heuristics raise the following question: when we are improving an existing partial spanning tree by addition of a (more advantageous)
partial subtree, which part of the original tree can be removed in order to eliminate the introduced cycles?  In the general case, the structure
which can be removed from a tree after the addition of another one (to find a spanning tree covering a given set of nodes) is a forest. We
give a reduction method of the arising problem as well as the algorithmic solution to identify the maximum removal forest ( 4 ).
Recently we are working on the limite of the approximation ratio of our propositions.

Multicast Communication with QoS
In certain cases analysis of multicast routing in packet switched communication networks when there is no available exact link state
information in the network is needed. From this analysis we develop novel algorithms for QoS routing. The multicast routing with
bandwidth and cost requirement in the case of incomplete information is reduced to a deterministic Steiner tree problem.  The simulation results show the performance which can be achieved by implementing the methods in QOSPF environment ( 5 ). A fast algorithm is proposed for the end-to-end delay problem in ( 6 ).

Multicast Communication in All Optical WDM Networks
The multicast communication becomes a particular and interesting problem in all optical WDM networks. Beyond the wavelength
conversion problem,  an other phenomena constrains the multicast diffusion: some switches can not split the  messages. The partial
spanning tree construction can be out of consideration for the splitting capability of nodes.

Fast Re-routing
Actually, OSPF convergence time is about 1 minute. However, this convergence time is not low enough to be tolerated by real time
applications. We propose complementary mechanism to OSPF which reduces the protocol convergence time by calculating in advance for
each possible destination in the network an alternate backup path which is disjoint from the primary used path. The TDSP "Two Disjoint
Shortest Paths" algorithm that we had proposed calculates two disjoint paths in one pass with a complexity of O(n2) (cf.  7  and  8 ).

Earlier activities

Optimization of Stochastic Inventory Systems
The application of stochastic approximation has been proposed in order to work out the optimization of inventory control problems when
the inventory manager does not have a priori knowledge about the random variables of the system. Multiple models was proposed in my thesis work( 9 ).

Neural Networks
Neural networks are good general function approximators.  Often the learning of neural networks is realized with the help of well known "back propagation" method.  The acceleration of the learning process with a second learning method is possible ( 10 ).


Publications
 

Teaching at the Computer Science Department of Insa