[calibrating a CA model with a genetic algorithm] - A New Kind of Science: The NKS Forum

Pages:1

# calibrating a CA model with a genetic algorithm

Posted by: Jason Cawley

I thought some here might be interested in a recent geological paper from a team in Italy that used the genetic algorithm technique to explore a space of CA models, to fit their empirical data. The paper appeared in the journal Geomorphology, March 2005 edition.

Applying Genetic Algorithms for Calibrating a Hexagonal Cellular Automata Model for the Simulation of Debris Flows Characterised by Strong Inertial Effects

Di Gregorio, S. and D'Ambrosio, D. and Iovine, G.

Abstact - In modelling complex a-centric phenomena which evolve through local interactions within a discrete time-space, cellular automata (CA) represent a valid alternative to standard solution methods based on differential equations. Flow-type phenomena (such as lava flows, pyroclastic flows, earth flows, and debris flows) can be viewed as a-centric dynamical systems, and they can therefore be properly investigated in CA terms. SCIDDICA S-4a is the last release of a two-dimensional hexagonal CA model for simulating debris flows characterised by strong inertial effects. S-4. has been obtained by progressively enriching an initial simplified model, originally derived for simulating very simple cases of slow-moving flow-type landslides. Using an empirical strategy, in S-4a the inertial character of the flowing mass is translated into CA terms by means of local rules. In particular, in the transition function of the model, the distribution of landslide debris among the cells is obtained through a double cycle of computation. In the first phase, the inertial character of the landslide debris is taken into account by considering indicators of momentum. In the second phase, any remaining debris in the central cell is distributed among the adjacent cells, according to the principle of maximum possible equilibrium. The complexities of the model and of the phenomena to be simulated suggested the need for an automated technique of evaluation for the determination of the best set of global parameters. Accordingly, the model is calibrated using a genetic algorithm and by considering the May 1998 Curti-Sarno (Southern Italy) debris flow. The boundaries of the area affected by the debris flow are simulated well with the model. Errors computed by comparing the simulations with the mapped areal extent of the actual landslide are smaller than those previously obtained without genetic algorithms. As the experiments have been realised in a sequential computing environment, they could be improved by adopting a parallel environment, which allows the performance of a great number of tests in reasonable times.

Posted by: Jason Cawley

As the authors themselves acknowledge, exhaustive search might be preferable when it is feasible, for this sort of thing. But they got surprisingly decent results exploring their moderately large parameter space with the GA scheme. The observed convergence of the GA is only partial - it does not find perfect fits - but pretty much monotonic and reasonably fast. Understand, they built the CA model from direct physical principles. The GA is just used to find what values for its variables seem to give the closest fit to a real landslide they wanted to re-simulate.