Benchmarks

The size of the landscape along with the number of traps in it determine the total processing power required for the optimization of the traps’ placements.

For our tests, we ran a regular square-grid landscape with an increasing number of points and traps. We ran this code with the base DEAP implementation of a genetic algorithm with Gaussian mutation, blend crossover and tournament selection. The algorithm was run for 1000 generations (and five repetitions per scenario) in a server with 88 cores, 3GHz processors (two 22 processor sockets) and 250GB RAM.

The scaling on landscape size follows this behavior:

_images/bench_PointsVTime.jpg

Whereas the scaling on traps number adheres to the following:

_images/bench_TrapsVTime.jpg

While these times are hardware-dependent (and we have a dedicated server to run our tests with tons of memory), the shape of the response to landscape sizes should help inform the decision on land size/traps number to be run in a general-use computer.

The code used for these benchmarks can be found in this link