Orateur
Prof.
Pascal Ballet
(Université de Bretagne Occidentale)
Description
Programming multiagent systems for biological systems close to
cellular scales require significant computational capabilities for
different reasons: the studied systems include many interacting
entities (usually more than 10^5), they have very different sizes
(from 10^-8 for the macromolecules to 10^-4 for small organisms) and
their behaviors are frequently complex. For example a cell changes its
shape, migrates, orientates, divides, adheres, communicate by direct
contact or at distance with others cells, etc.
The speed of microprocessors, in term of clock frequency, reaching its
limits (around 4GHz), major manufacturers have chosen to develop
multicore processors. This evolution is seen in Central Processor
Units (CPU) but also in Graphic Processor Units (GPU). GPU contains
stream processors allowing the parallel treatment of data (Single
Instruction Multiple Data). Moreover, with the incoming of the openCL
language, it is now possible to use he power of stream processing on
different hardware (mother board and graphic cards for instance).
Thanks to this, we have developed reactive multiagent algorithms and a
simple software architecture to simulate multicellular phenomena with
many interacting entities (more than 10^5). We show different
simulations like a simple random walk, a Belousov-Zhabotinsky like
reaction, a prey-predator system and a multicellular morphogenesis
system. The examples are detailed and their efficiencies and drawbacks
are discussed.