Practical Genetic Algorithm Settings for Amesim Optimization

Amesim's built-in optimization (and its coupling to Simcenter HEEDS for multi-objective work) exposes genetic algorithm (GA) parameters that are easy to leave at defaults without understanding their practical impact. Volupe (a Siemens Amesim reseller/training partner) publishes concrete rule-of-thumb guidance that goes beyond the official documentation's parameter descriptions.

Population size

Rule of thumb: population size ≥ 4.5 × number of design parameters tends to give good results. This is a practical starting point for sizing a GA run when you don't yet have a feel for the design space — too small a population relative to parameter count risks premature convergence or missing good regions of the space entirely.

Reproduction ratio

Controls what fraction of each generation is replaced (vs. retained as "elite" parents) between iterations. Amesim's default is 80% (i.e., the best 20% are retained). The suggested effective range is 50%–85% for good results — going outside this range in either direction tends to hurt either exploration (too high a retention/too low a reproduction ratio) or convergence stability (too aggressive a reproduction ratio).

Number of generations

No universal number — it's fundamentally a compute-budget decision, since more generations cost more simulation runs. But there is a hard floor: it must be greater than 10 to get results that are actually meaningful/relevant. Fewer than that and the GA hasn't had enough iterations to demonstrate real convergence behavior.

Mutation probability & amplitude

Applied specifically to discrete parameters. Mutation amplitude is what helps the algorithm continue exploring the design space and avoid getting stuck prematurely converging on a local optimum — worth increasing deliberately if you suspect the GA is converging too fast to a mediocre result.

Seed value

The GA uses a pseudo-random seed. Using the same seed reproduces identical results run to run (useful for debugging/reproducibility), while different seeds explore different search paths through the design space (useful for confirming a result isn't an artifact of one particular random trajectory — i.e., run the optimization multiple times with different seeds if you need confidence in the optimum found).

Total simulation count formula

A useful sanity-check formula before committing to a long optimization run:

Total simulations = N × (1 − R) × G

where N = population size, R = reproduction ratio, G = number of generations. Use this up front to estimate wall-clock cost before launching a large sweep, especially when each individual simulation is expensive.

When to prefer GA over gradient-based optimization

GAs are better suited to multi-modal problems (design spaces with multiple local optima) because they don't converge prematurely to the nearest local minimum the way gradient-based methods can. The trade-off is that GAs require substantially more computational resources (many more simulation runs) than a gradient-based approach — so for well-behaved, single-optimum problems, a gradient method will typically get you to the answer much faster.

Parameter bounds when coupling to Simcenter HEEDS

A specific gotcha when driving an Amesim model from Simcenter HEEDS for multi-objective optimization: care must be taken defining the lower/upper limits of each parameter in HEEDS. If HEEDS's optimizer proposes a parameter value outside the range Amesim expects, the Amesim model simply will not be updated with that value for that run — silently invalidating that design point rather than throwing an obvious error. Always double check parameter bound consistency between the two tools before trusting optimization results.

Source: Simcenter Amesim's Built-In Optimization Features – Volupe, Multi-Objective Optimization – HEEDS & Amesim – Volupe