burlap.behavior.singleagent.learning.modellearning.modelplanners

Class VIModelLearningPlanner

java.lang.Object

burlap.behavior.singleagent.MDPSolver

burlap.behavior.singleagent.planning.stochastic.DynamicProgramming

burlap.behavior.singleagent.planning.stochastic.valueiteration.ValueIteration

burlap.behavior.singleagent.learning.modellearning.modelplanners.VIModelLearningPlanner

All Implemented Interfaces:

ModelLearningPlanner, MDPSolverInterface, Planner, QFunction, QProvider, ValueFunction

public class VIModelLearningPlanner
extends ValueIteration
implements ModelLearningPlanner

A model learning interface wrapper to VI that causes VI to be performed every time the model is updated or whenever a novel state is seen that was not previously expected to be reachable. When the model changes, planning is always performed from the initial state of an episode as well as the last changed episode

Author:

James MacGlashan

Nested Class Summary

Nested classes/interfaces inherited from interface burlap.behavior.valuefunction.QProvider

QProvider.Helper

Field Summary

Fields

Modifier and Type	Field and Description
protected State	initialState The last initial state of an episode
protected Policy	modelPolicy The greedy policy that results from VI
protected java.util.Set<HashableState>	observedStates States the agent has observed during learning.

Fields inherited from class burlap.behavior.singleagent.planning.stochastic.valueiteration.ValueIteration

foundReachableStates, hasRunVI, maxDelta, maxIterations, stopReachabilityFromTerminalStates

Fields inherited from class burlap.behavior.singleagent.planning.stochastic.DynamicProgramming

operator, valueFunction, valueInitializer

Fields inherited from class burlap.behavior.singleagent.MDPSolver

actionTypes, debugCode, domain, gamma, hashingFactory, model, usingOptionModel

Constructor Summary

Constructors

Constructor and Description
VIModelLearningPlanner(SADomain domain, FullModel model, double gamma, HashableStateFactory hashingFactory, double maxDelta, int maxIterations) Initializes

Method Summary

Modifier and Type	Method and Description
void	initializePlannerIn(State s) This is method is expected to be called at the beginning of any new learning episode.
void	modelChanged(State changedState) Tells the valueFunction that the model has changed and that it will need to replan accordingly
Policy	modelPlannedPolicy() Returns a policy encoding the planner's results.
protected void	rerunVI() Reruns VI on the new updated model.

Methods inherited from class burlap.behavior.singleagent.planning.stochastic.valueiteration.ValueIteration

performReachabilityFrom, planFromState, recomputeReachableStates, resetSolver, runVI, toggleReachabiltiyTerminalStatePruning

Methods inherited from class burlap.behavior.singleagent.planning.stochastic.DynamicProgramming

computeQ, DPPInit, getAllStates, getCopyOfValueFunction, getDefaultValue, getModel, getOperator, getValueFunctionInitialization, hasComputedValueFor, loadValueTable, performBellmanUpdateOn, performBellmanUpdateOn, performFixedPolicyBellmanUpdateOn, performFixedPolicyBellmanUpdateOn, qValue, qValues, setOperator, setValueFunctionInitialization, value, value, writeValueTable

Methods inherited from class burlap.behavior.singleagent.MDPSolver

addActionType, applicableActions, getActionTypes, getDebugCode, getDomain, getGamma, getHashingFactory, setActionTypes, setDebugCode, setDomain, setGamma, setHashingFactory, setModel, solverInit, stateHash, toggleDebugPrinting

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Methods inherited from interface burlap.behavior.singleagent.planning.Planner

planFromState

Methods inherited from interface burlap.behavior.singleagent.MDPSolverInterface

addActionType, getActionTypes, getDebugCode, getDomain, getGamma, getHashingFactory, getModel, resetSolver, setActionTypes, setDebugCode, setDomain, setGamma, setHashingFactory, setModel, solverInit, toggleDebugPrinting

Field Detail

observedStates

protected java.util.Set<HashableState> observedStates

States the agent has observed during learning.

modelPolicy

protected Policy modelPolicy

The greedy policy that results from VI

initialState

protected State initialState

The last initial state of an episode

Constructor Detail

VIModelLearningPlanner

public VIModelLearningPlanner(SADomain domain,
                              FullModel model,
                              double gamma,
                              HashableStateFactory hashingFactory,
                              double maxDelta,
                              int maxIterations)

Initializes

Parameters:

domain - model domain

model - the learned model to use for planning

gamma - discount factor

hashingFactory - the hashing factory

maxDelta - max value function delta in VI

maxIterations - max iterations of VI

Method Detail

initializePlannerIn

public void initializePlannerIn(State s)

Description copied from interface: ModelLearningPlanner

This is method is expected to be called at the beginning of any new learning episode. This may be useful for planning algorithms that do not solve the policy for every state since new episodes may starts in states the planning algorithm had not previously considered. before a learning episode begins.

Specified by:

initializePlannerIn in interface ModelLearningPlanner

Parameters:

s - the input state

modelChanged

public void modelChanged(State changedState)

Description copied from interface: ModelLearningPlanner

Tells the valueFunction that the model has changed and that it will need to replan accordingly

Specified by:

modelChanged in interface ModelLearningPlanner

Parameters:

changedState - the source state that caused a change in the model.

modelPlannedPolicy

public Policy modelPlannedPolicy()

Description copied from interface: ModelLearningPlanner

Returns a policy encoding the planner's results.

Specified by:

modelPlannedPolicy in interface ModelLearningPlanner

Returns:

a policy object

rerunVI

protected void rerunVI()

Reruns VI on the new updated model. It will force VI to consider all states the agent has ever previously observed, even though not all may be connected by the current unknown transition model.