New section in GSTAM docs!

[Dellaert, Frank]

Thanks to Porter Zach’s efforts (and Gemini !) we have another section ready in the GTSAM docs: inference.

 

Of note is the section about Ordering: it hints at the difference between COLAMD and METIS. The former is the default, but the latter can be dramatically faster, depending on your problem’s sparsity structure. As always, you can open the notebook in Colab and experiment with the sample code.

 

Porter is working on automatic cross-referencing between doc pages.

 

Here is a handy link tree to all the new docs, please do create an issue if you detect inaccuracies:

 

The inference module provides the foundational classes and algorithms for probabilistic graphical models in GTSAM, focusing on variable elimination and the resulting structures like Bayes Nets and Bayes Trees.

Core Concepts¶

• Key: Base type (uint64_t) for uniquely identifying variables.
• Symbol: A Key type encoding a character and an index (e.g., x0).
• LabeledSymbol: A Symbol variant with an additional label character, useful for multi-robot scenarios (e.g., xA0).
• EdgeKey: A Key type encoding a pair of 32-bit integers.
• Factor: Abstract base class for all factors (relationships between variables).
• FactorGraph: Base class representing a collection of factors.
• Conditional: Abstract base class for conditional distributions/densities resulting from elimination ( P(Frontals∣Parents)P(Frontals∣Parents) ).

Elimination Algorithms & Control¶

• Ordering: Specifies the order in which variables are eliminated, crucial for efficiency.
• VariableIndex: Maps variables to the factors they appear in, used for efficient elimination ordering and construction.
• EliminateableFactorGraph: A mixin class providing eliminateSequential and eliminateMultifrontal methods to concrete factor graph types (like GaussianFactorGraph, SymbolicFactorGraph).

Elimination Results & Structures¶

• BayesNet: Represents the result of sequential variable elimination as a directed acyclic graph (DAG) of conditionals.
• EliminationTree: Tree structure representing the dependencies and computations during sequential elimination.
• ClusterTree: Base class for tree structures where nodes are clusters of factors (e.g., JunctionTree).
• JunctionTree: A cluster tree representing the cliques formed during multifrontal elimination, holding the factors before they are eliminated into conditionals.
• BayesTreeCliqueBase: Abstract base class for the nodes (cliques) within a BayesTree.
• BayesTree: Represents the result of multifrontal variable elimination as a tree of cliques, where each clique contains the conditional P(Frontals∣Separator)P(Frontals∣Separator).

Incremental Inference¶

• ISAM: Incremental Smoothing and Mapping algorithm based on updating a BayesTree (original version, often superseded by ISAM2 in nonlinear).

Visualization¶

• DotWriter: Helper class to customize the generation of Graphviz .dot files for visualizing graphs and trees.