Tè Darwiniano 

In occasione del undicesimo incontro della Darwinian Tea Society of Rome (!), i presenti (Raffaele Calabretta, Daniele Denaro, Andrea Di Ferdinando, Davide Marocco, Stefano Nolfi, Luigi Pagliarini e Domenico Parisi), hanno salutato
Ketil Eiane, biologo marino dell'Università di Bergen e post-doctoral fellow al GRAL, in partenza da Roma per trasferirsi all'Università di Spitsbergen (nord della Norvegia) dove insegnerà Ecology and Evolution.

Ketil ci ha parlato del suo lavoro svolto al Gral in collaborazione con Domenico Parisi:
 

 
Zooplankton are small animals living in the free water masses. They drift horizontally with the water currents, but are capable of migrating vertically. Such behavior generally occur on a daily basis: they spend night close to the surface and migrate to depths of up to 200 m at daytime.

The biological significance of this behavior is understood in terms a trade off between the need for food consumption (food is abundant in surface waters) and the risk of being killed by a predator (fish being most effective predators on zooplankton in well lit surroundings, i.e. in surface waters at daytime).

We have applied a "stimuli – response" approach to model this behavioral pattern. In our model behavior is controlled by a neural network (NN) forced with sensory information we know zooplankton capable of perceiving (light level, food concentration, presence of predators). In addition we included in our simulations a modified NN where also a simple internal clock forms an additional 4^th input parameter. The output generated by the model is the decision to move vertically in the water column. Weights in the NN is optimized by using a genetic algorithm (GA) where populations of 1000 individuals are introduced into an artificial environment resembling a water column. After running the model for 5 days (with a time step of 2 min) the 50 best individuals are selected for forming the gene-pool for the next generation. During this procedure both random point-mutations and crossover is applied.

From a pilot study we found that a NN with 4 nodes in the hidden layer performed best for this task. The model performed reasonably well both under a deterministic environment, and proved to be fairly robust to increasing levels of stochasticity in the environmental parameters. The effect of including an internal timer to the network did not affect the performance to any significant degree when tested to an environment close in day length to what it had been adapted for. When testing for robustness to different day-lengths (mid-winter vs. mid-summer) however, we found that the NN without the internal clock performed reasonably well irrespective of selection regime. The NN with an internal clock however, fails when tested for a different day-length regime than adapted for.

Last updated: lunedì 09 luglio 2001
Torna all'Home Page del Tè Darwiniano