jpt.inference.mpe

Classes

`MPEState`
`BnBNode`
`BnBSolution`
`MPESolver`	Solver for k-MPE inference about n independent variables

Module Contents

class jpt.inference.mpe.MPEState(assignment: jpt.variables.VariableMap)

assignment

latest_variable = None

copy()

assign(variable: Variable or str, value: Any) → MPEState

__repr__()

class jpt.inference.mpe.BnBNode(solver, state, actual_cost)

solver: MPESolver

old_state: MPEState

variable: jpt.variables.Variable = None

values: Iterator[Any]

actual_cost

step_cost = 0

state: MPEState = None

_generate_next()

pop()

property cost

property free_variables

class jpt.inference.mpe.BnBSolution(state: MPEState, cost: float)

state

cost

class jpt.inference.mpe.MPESolver(distributions: jpt.variables.VariableMap, likelihood_divisor: float = None)

Solver for k-MPE inference about n independent variables

This algorithm iteratively constructs all k most probable explanations in descending order by performing a branch-and-bound search in the space of atomic areas of the respective distributions. In constrast to classic BnB search, we do not throw away the pruned states but save them in a priority queue, where we can continue the search for the subsequent 2nd, 3rd, …, k-th best solution.

distributions: jpt.variables.VariableMap

domains

constraints

lb

likelihood(solution: BnBSolution)

is_goal_state(state: MPEState) → bool

variable_order(state: MPEState = None) → Iterator[jpt.variables.Variable]

Parameters:: state – The state to start from
Returns:: An iterator over free variables for a state sorted by number of different possible states (lowest amount of values first).

solve(k: int = 0) → Iterator[List[jpt.variables.ValueAssignment]]

Generate k MPE states with decreasing probability.

Returns: