Parameter optimization

optimize_simulation_parameters() pairs a single simulation task with a scipy-based Nelder-Mead solver. At each iteration the optimizer picks new parameter values, runs the simulation, reads the target scalars, and minimizes the difference from the expected values. No gradients needed -- suitable for stochastic simulations.

Upstream reference: https://github.com/omnetpp/opp_repl/blob/main/doc/parameter_optimization.md

Requires the optimize extra (scipy, optimparallel):

pip install "opp_repl[optimize]"

Signature (keyword-only)

optimize_simulation_parameters( simulation_task, # result of get_simulation_task(...) expected_result_names=[...], # scalar names to hit expected_result_values=[...], # target values fixed_parameter_names=[...], # optional -- held constant fixed_parameter_values=[...], fixed_parameter_assignments=[...], # INI-style path fixed_parameter_units=[...], parameter_names=[...], # what to vary parameter_assignments=[...], # INI-style path for each parameter_units=[...], # format string(s); see below initial_values=[...], min_values=[...], max_values=[...], )

Unit format strings

• Plain unit: "m", "Mbps" -- appended to the numeric value.
• Distribution wrapper: "exponential({0}s)" -- preserves an

iaTime distribution while overriding its parameter.

• Any .format(value)-compatible string.

This matters because many OMNeT++ parameters are declared volatile and assigned distributions in the INI file. The unit wrapper keeps the distribution intact.

Example 1 — slotted ALOHA channel utilization

Maximize channel utilization in slotted ALOHA. Theoretical peak is 1/e ≈ 0.368 at iaTime ≈ 1.87 s. From overloaded start (0.5 s), converges in ~40 evaluations:

optimize_simulation_parameters( get_simulation_task(config_filter="SlottedAloha1", sim_time_limit="10min"), expected_result_names=["channelUtilization:last"], expected_result_values=[0.368], fixed_parameter_names=[], fixed_parameter_values=[], fixed_parameter_assignments=[], fixed_parameter_units=[], parameter_names=["iaTime"], parameter_assignments=["Aloha.host[*].iaTime"], parameter_units=["exponential({0}s)"], initial_values=[0.5], min_values=[0.1], max_values=[20])

Output ends with:

Best: {'iaTime': 1.873} -> {'channelUtilization:last': 0.367}

Example 2 — INET WiFi range at 30 % PER

Find the distance where 54 Mbps WiFi reaches 30 % packet error. Converges to ~53.2 m in ~28 evaluations:

optimize_simulation_parameters( get_simulation_task(simulation_project=inet_project, working_directory_filter="showcases/wireless/errorrate", config_filter="General", run_number=0, sim_time_limit="1s"), expected_result_names=["packetErrorRate:vector"], expected_result_values=[0.3], fixed_parameter_names=["bitrate"], fixed_parameter_values=[54], fixed_parameter_assignments=["*.bitrate"], fixed_parameter_units=["Mbps"], parameter_names=["distance"], parameter_assignments=[".destinationHost.mobility.initialX"], parameter_units=["m"], initial_values=[50], min_values=[20], max_values=[100])

Pitfalls

• Stochastic noise dominates near the optimum. Lengthen

sim_time_limit or average several runs per evaluation (repeat count in the INI) when results get noisy.

• get_simulation_task() MUST match exactly one task. Ambiguous

filters raise -- add run_number=0 or tighter filters.

• Bounds and initial values should straddle the expected optimum.

Nelder-Mead can get stuck if the initial simplex is far from the feasible region.

• Each evaluation is a full simulation run; budget accordingly.
• Multi-dimensional optimization (several parameter_names) scales

roughly O(n²) in simplex size; stay ≤ 4-5 parameters for sanity.

See also

• opp-repl-running-simulations — get_simulation_task() semantics.
• opp-repl-tasks-and-results — reading result scalars.
• opp-repl-ssh-cluster — distribute parallel evaluations.