LevenbergMarquardtParams

LevenbergMarquardtParams combines the common nonlinear-optimizer settings with damping controls specific to Levenberg–Marquardt. It determines how the optimizer moves between Gauss–Newton and gradient-descent behavior.

Authors: Frank Dellaert, et al. (see THANKS for the full author list)

See LICENSE for the license information

try:
    import google.colab
    %pip install --quiet gtsam-develop
except ImportError:
    pass

import gtsam
import numpy as np
from gtsam.symbol_shorthand import L, V, X

Defaults and construction¶

The ordinary constructor uses GTSAM defaults. CeresDefaults() and LegacyDefaults() provide named parameter sets for compatibility with those conventions.

params = gtsam.LevenbergMarquardtParams.CeresDefaults()
print("maximum iterations:", params.getMaxIterations())
print("initial lambda:", params.getlambdaInitial())
print("lambda factor:", params.getlambdaFactor())

Damping controls¶

lambdaInitial, lower/upper bounds, and lambdaFactor control damping adaptation. setDiagonalDamping(True) scales damping by the Hessian diagonal; setUseFixedLambdaFactor() selects fixed multiplicative updates.

params.setlambdaInitial(1e-3)
params.setlambdaLowerBound(1e-8)
params.setlambdaUpperBound(1e5)
params.setlambdaFactor(5.0)
params.setDiagonalDamping(True)
params.setVerbosityLM("SILENT")

Using the parameters¶

Common inherited methods select convergence tolerances, iteration limits, ordering, and linear solver. The optimizer copies the parameter object at construction.

graph = gtsam.NonlinearFactorGraph()
model = gtsam.noiseModel.Diagonal.Sigmas(np.array([0.1, 0.1, 0.05]))
graph.add(gtsam.PriorFactorPose2(X(0), gtsam.Pose2(1.0, 2.0, 0.3), model))
initial = gtsam.Values()
initial.insert(X(0), gtsam.Pose2(0.0, 0.0, 0.0))

optimizer = gtsam.LevenbergMarquardtOptimizer(graph, initial, params)
result = optimizer.optimize()
print("optimized pose:", result.atPose2(X(0)))
assert graph.error(result) < graph.error(initial)

Diagnosing damping behavior¶

If lambda repeatedly reaches its upper bound, inspect scaling, initialization, and factor Jacobians instead of simply increasing the bound. If convergence is smooth but slow, compare diagonal damping and the named default sets, while keeping the same stopping tolerances for a fair comparison.

Source¶

LevenbergMarquardtParams.h

AI assistance caveat¶

AI was used to help draft this documentation, and inaccuracies could be present.