filtering
Trait for creating and running Belief Propagation within Factored Frontier.
Abstract class that runs the Factored Frontier algorithm.
Abstract class that runs the Factored Frontier algorithm. Like a particle filter, the algorithm is supplied with models representing initial and static universes, as well as a universe transition function.
At each time step, the algorithm copies the marginal probabilities for each named element to a new dummy universe. This dummy universe is then supplied to the transition function.
The general class of filtering algorithms.
The general class of filtering algorithms. A filtering algorithm is provided with an initial model, represented by a universe encoding the probability distribution over the initial state, and a transition model, which maps a state to a universe encoding the probability distribution over the new state. An implementation of Filtering must implement the advanceTime, computeCurrentDistribution, and computeCurrentExpectation methods.
Querying and asserting evidence to a filtering algorithm are done using references. This is because references are stable over time, while the particular elements they refer to are not.
One-time filtering algorithms.
One-time filtering algorithms. An implementation of OneTimeFiltering must implement the advanceTime, computeCurrentDistribution, and computeCurrentExpectation methods.
A one-time particle filter.
A parallel version of Filtering.
A parallel version of Filtering. Specifically a version of OneTimeFiltering, since that's the only target for parallelization right now.
A parallel one-time particle filter.
A parallel one-time particle filter. Distributes the work of generating particles at each time step over a specified number of threads. After generating the particles, they are recombined before re-sampling occurs. Instead of accepting initial and static universes as input, this method accepts functions that return universes. This is because each thread needs its own set of universes to work on. It is important that any elements created within those functions are explicitly assigned to the returned universe, not the implicit default universe.
An abstract class of particle filters.
An abstract class of particle filters. A particle filter is provided with three models: a static model, containing a universe defining a distribution over static elements that do not change over time; an initial model, containing a universe defining a distribution over the initial state of time-varying elements; and a transition model, which is a function from the previous universe to a new universe. defining the way the distribution over the new state of the time-varying variables depends on their values in the previous state. The fourth argument to the particle filter is the number of particles to use at each time step.
The particle filter works in an online fashion. At each point in time, it maintains its current beliefs about the state of the system as a set of representative states. advanceTime is used to move forward one time step. The particle filter updates its beliefs in light of the new evidence.
A record of the values of a set of named elements at a particular point in time.
A state contains snapshots of the static universe and the dynamic universe, representing a configuration of the static and dynamic variables in the system.
A class representing a single window in time, with a current universe, a previous universe, and a static universe.
A parallel implementation of a OneTimeParticleFilter.
Trait for creating and running Belief Propagation within Factored Frontier. BP runs as either a one time or anytime algorithm.