A discrete design space
OPT is the principal Sim::OPT optimization and parametric-exploration program. It coordinates the definition of a discrete multidimensional design space, the organization of variables into search blocks, model morphing, simulation, result retrieval, reporting and the progression of the exploration.
A problem is described by a Perl configuration file. Search cases and blocks are defined through @sweeps; the available levels of the design variables are described through the variable-level structures, and a central or root instance can be specified as the starting point of the exploration.
The variables included in different blocks may overlap. This is deliberate: the decomposition is part of the search strategy rather than a rigid partition of the problem.
Model transformation and simulation
OPT can operate on simulation models that receive text-based input and emit numerical or textual results. Sim::OPT includes general mechanisms for manipulating model files and also ESP-r-specific functions for building geometry, shading, networks, controls and result retrieval.
Constraint-propagation configurations can be used when a change to one part of a model must be propagated to other geometric or non-geometric quantities. The resulting design instance is therefore not merely a vector of numbers: it can correspond to a fully transformed simulation model.
Search structures
Sim::OPT supports coordinate and block searches, sequential and parallel forms of progression, star searches, factorial exploration of pre-simulated or metamodel data, and face-centred composite designs.
Sequential block searches correspond to an inexact Gauss-Seidel logic in which the result of one block can affect the starting state of the next. Parallel block exploration corresponds to an inexact Jacobi logic in which several block results can be generated from a common state.
Star searches can use the historical diagonal subdivision or an undiagonal deterministic subdivision intended to spread points more evenly through a discrete lattice. Explicitly supplied star positions take precedence.
Precomputed and reconstructed data
OPT does not require every objective value to be obtained by a simulation launched during the current run. It can operate on previously calculated results, and selected blocks can be supported by reconstructed or metamodelled data.
This allows the same search machinery to be used both for direct simulation-supported optimization and for exploration of an existing discrete dataset.