Release notes

Most important changes in Indago releases are listed here. Breaking changes are indicated by the sign.

Indago 0.6.0

Released on 2026-04-20

New methods: Heap-Based Optimizer (HBO), Controlled Random Search (CRS)
New samplers for generating initial population positions (Optimizer.sampler): Sobol ('sobol') and LatinHypercube ('lhs')

IndagoBench25 results overview and corresponding guidelines added to documentation
Template file for new methods included in the methods library (_new_method_template.py)
Anakatabatic PSO variants are now using scipy.interpolate.make_interp_spline instead of legacy function interp1d
Literature reference for MSGD provided in documentation
LSHADE set as the default DE variant (performs much better than SHADE)
Optimizer.scatter_method renamed to Optimizer.sampler

Fixed the problem of EFO crashing when using small pop_size values
Using custom anakatabatic PSO without akb_model parameter now works without errors
Fixed the error appearing when logging messages for some methods

Released on 2025-11-16

MSGS method has been completely rewritten, improved and renamed to Multi-Scale Gradient Descent (MSGD)
MSGS removed from Indago

Solved bug in Optimizer.plot_history (negative linthresh in symlog was triggered if c<=0 in all iterations)

Released on 2024-12-16

Adding ESC (Escape Algorithm) as unsupported optimizer (not documented nor tested)

Candidate.copy() now copies uniqe_str attribute. There was a problem when accessing uniqe_str in assigned Optimizer.post_iteration_processing function.

Released on 2024-11-02

Indago methods (most of them anyway) can now solve unbounded (or mixed) problems. Bounds (Optimizer.lb, Optimizer.ub) can be skipped, or defined (partially or fully) by +/-np.inf or np.nan values. Indago will replace “open” bounds with +/-1e100. When solving an unbound problem Optimizer.X0 must be provided
Implementation of Status enum class for optimization status tracking (accessible via Optimizer.status attribute)

All methods are now significantly faster due to code optimizations in Candidate class
Optimizer.optimize() now returns a Candidate instance regardless of the method being used
A multilevel candidate comparison operator used in Indago methods is now clearly explained in the documentation
Behavior of indago.unconstrain() is now better explained in the documentation
Better formatting of elapsed time in dashboard monitoring output

Fixed an issue with occurrences of NaN fitness values in MSGS
Fixed an issue with MSGS exiting the search space bounds
Optimizer.plot_history() is further improved to avoid division-by-zero warnings.
Fixed MSGS status and log message for xtol stopping criterion

Released on 2024-07-16

Color tweaking and adding method name in history/convergence plot
Optimization runs can now be resumed and restarting optimizations should not produce problems
Some code refactoring
Better literature referencing in documentation
New default population (pop_init) for DE: max(30, 5 * dimensions)
Reduced computational cost of BA

Released on 2024-01-12

HSIW inertia technique for PSO
FullyEmployed variant of ABC
scatter_method optimizer parameter, with ‘halton’ option for using Halton sequence for initializing design vectors
New method: Random Search (RS)
Optimizer.plot_history() is improved to produce better and fancier plots.
Export of all evaluated candidate solutions (Optimizer.evals_db parameter)
New utility functions: unconstrain() and read_evals_db()
New method: Grey Wolf Optimization (GWO)
Checking for unknown (illegal) optimizer attributes
Some important optimizer parameters can be given as evaluation_function attributes (namely dimensions, lb, ub, objectives, constraints, objective_weights, objective_labels, constraint_labels)
Optimizer.status providing textual information on the optimizer state
CandidateState.is_feasible() for checking whether the solution is feasible (i.e. whether it satisfies all constraints)
Optimizer.copy() produces a (deep) copy of an Optimizer object

Cleaner log messages
Better documentation
Completely rewritten ABC (significantly different behavior expected)
New (shorter and more consistent) param names in almost all methods
NelderMead method now has an abbreviation (NM), just like all other methods
CEC 2014 benchmark test is removed (we are working on a much more comprehensive benchmark test suite)
New defaults in minimize_exhaustive()
More consistent error reporting style
New (better) default param values for all methods
Better evaluation error handling
number_of_processes parameter is now called processes
Updated requirements (most importantly python>=3.9)
Optimizer.plot_history() can skip subplots if needed. Useful for turning off dysfunctional subplots with too many plot lines.

target_fitness stopping criterion is not ignored any more
Fixed stopping criteria overshooting defined targets
Fixed memory leak due to circular referencing in Optimizer and CandidateState objects
Fixed multiple problems in MSGS
Fixed some evaluation retrying problems
Rich-based features adapted to a new version of rich