burlap.behavior.stochasticgame.mavaluefunction.vfplanners

Class MAValueIteration

java.lang.Object

burlap.behavior.stochasticgame.mavaluefunction.MAValueFunctionPlanner

burlap.behavior.stochasticgame.mavaluefunction.vfplanners.MAValueIteration

All Implemented Interfaces:

MultiAgentQSourceProvider

public class MAValueIteration
extends MAValueFunctionPlanner

A class for performing multi-agent value iteration. This class extends the MAValueFunctionPlanner class to provide value iteration-like value function estimation. When an input state is provided via the planFromState(State) method, if the state has already been seen and planned for, then nothing happens. If the state has never been seen before, then a state reachiability analysis is first performed in which all states possibly reachable from the input state are found. Then Value iteration proceeds for all states that have been found in the past. The runVI() method can also be called directly to force value iteration to be performed on all states that have been previously found, but the state reachability must have been performed at least once before to seed the state space. State reachability can be performed manually by calling the performStateReachabilityFrom(State) method.

Value iteration will continue until either the maximum change in Q-value is less than some user provided threshold or until a max number of iterations have passed.

Author:

James MacGlashan

Nested Class Summary

Nested classes/interfaces inherited from class burlap.behavior.stochasticgame.mavaluefunction.MAValueFunctionPlanner

MAValueFunctionPlanner.BackupBasedQSource, MAValueFunctionPlanner.JointActionTransitions

Field Summary

Fields

Modifier and Type	Field and Description
protected int	debugCode The debug code used for printing VI progress.
protected double	maxDelta The threshold that will cause VI to terminate when the max change in Q-value for is less than it
protected int	maxIterations The maximum allowable number of iterations until VI termination
protected java.util.Set<StateHashTuple>	states The set of states that have been found

Fields inherited from class burlap.behavior.stochasticgame.mavaluefunction.MAValueFunctionPlanner

agentDefinitions, backupOperator, discount, domain, hashingFactory, jointActionModel, jointReward, planningStarted, qSources, terminalFunction, vInit

Constructor Summary

Constructors

Constructor and Description
MAValueIteration(SGDomain domain, JointActionModel jointActionModel, JointReward jointReward, TerminalFunction terminalFunction, double discount, StateHashFactory hashingFactory, double qInit, SGBackupOperator backupOperator, double maxDelta, int maxIterations) Initializes.
MAValueIteration(SGDomain domain, JointActionModel jointActionModel, JointReward jointReward, TerminalFunction terminalFunction, double discount, StateHashFactory hashingFactory, ValueFunctionInitialization qInit, SGBackupOperator backupOperator, double maxDelta, int maxIterations) Initializes.
MAValueIteration(SGDomain domain, java.util.Map<java.lang.String,AgentType> agentDefinitions, JointActionModel jointActionModel, JointReward jointReward, TerminalFunction terminalFunction, double discount, StateHashFactory hashingFactory, double vInit, SGBackupOperator backupOperator, double maxDelta, int maxIterations) Initializes.
MAValueIteration(SGDomain domain, java.util.Map<java.lang.String,AgentType> agentDefinitions, JointActionModel jointActionModel, JointReward jointReward, TerminalFunction terminalFunction, double discount, StateHashFactory hashingFactory, ValueFunctionInitialization vInit, SGBackupOperator backupOperator, double maxDelta, int maxIterations) Initializes.

Method Summary

Methods

Modifier and Type	Method and Description
boolean	performStateReachabilityFrom(State s) Finds and stores all states that are reachable from input state s.
void	planFromState(State s) Calling this method causes planning to be performed from State s.
void	runVI() Runs Value Iteration over the set of states that have been discovered.

Methods inherited from class burlap.behavior.stochasticgame.mavaluefunction.MAValueFunctionPlanner

backupAllValueFunctions, getQSources, hasStartedPlanning, initMAVF, setAgentDefinitions

Methods inherited from class java.lang.Object

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

Field Detail

states

protected java.util.Set<StateHashTuple> states

The set of states that have been found

maxDelta

protected double maxDelta

The threshold that will cause VI to terminate when the max change in Q-value for is less than it

maxIterations

protected int maxIterations

The maximum allowable number of iterations until VI termination

debugCode

protected int debugCode

The debug code used for printing VI progress.

Constructor Detail

MAValueIteration

public MAValueIteration(SGDomain domain,
                JointActionModel jointActionModel,
                JointReward jointReward,
                TerminalFunction terminalFunction,
                double discount,
                StateHashFactory hashingFactory,
                double qInit,
                SGBackupOperator backupOperator,
                double maxDelta,
                int maxIterations)

Initializes.

Parameters:

domain - the domain in which to perform planing

jointActionModel - the joint action model

jointReward - the joint reward function

terminalFunction - the terminal state function

discount - the discount

hashingFactory - the hashing factory to use for storing states

qInit - the default Q-value to initialize all values to

backupOperator - the backup operator that defines the solution concept being solved

maxDelta - the threshold that causes VI to terminate when the max Q-value change is less than it

maxIterations - the maximum number of iterations allowed

MAValueIteration

public MAValueIteration(SGDomain domain,
                JointActionModel jointActionModel,
                JointReward jointReward,
                TerminalFunction terminalFunction,
                double discount,
                StateHashFactory hashingFactory,
                ValueFunctionInitialization qInit,
                SGBackupOperator backupOperator,
                double maxDelta,
                int maxIterations)

Initializes.

Parameters:

domain - the domain in which to perform planing

jointActionModel - the joint action model

jointReward - the joint reward function

terminalFunction - the terminal state function

discount - the discount

hashingFactory - the hashing factory to use for storing states

qInit - the q-value initialization function to use.

backupOperator - the backup operator that defines the solution concept being solved

maxDelta - the threshold that causes VI to terminate when the max Q-value change is less than it

maxIterations - the maximum number of iterations allowed

MAValueIteration

public MAValueIteration(SGDomain domain,
                java.util.Map<java.lang.String,AgentType> agentDefinitions,
                JointActionModel jointActionModel,
                JointReward jointReward,
                TerminalFunction terminalFunction,
                double discount,
                StateHashFactory hashingFactory,
                double vInit,
                SGBackupOperator backupOperator,
                double maxDelta,
                int maxIterations)

Initializes.

Parameters:

domain - the domain in which to perform planing

agentDefinitions - the agents involved in the planning problem

jointActionModel - the joint action model

jointReward - the joint reward function

terminalFunction - the terminal state function

discount - the discount

hashingFactory - the hashing factory to use for storing states

vInit - the default value to initialize all state values to

backupOperator - the backup operator that defines the solution concept being solved

maxDelta - the threshold that causes VI to terminate when the max Q-value change is less than it

maxIterations - the maximum number of iterations allowed

MAValueIteration

public MAValueIteration(SGDomain domain,
                java.util.Map<java.lang.String,AgentType> agentDefinitions,
                JointActionModel jointActionModel,
                JointReward jointReward,
                TerminalFunction terminalFunction,
                double discount,
                StateHashFactory hashingFactory,
                ValueFunctionInitialization vInit,
                SGBackupOperator backupOperator,
                double maxDelta,
                int maxIterations)

Initializes.

Parameters:

domain - the domain in which to perform planing

agentDefinitions - the agents involved in the planning problem

jointActionModel - the joint action model

jointReward - the joint reward function

terminalFunction - the terminal state function

discount - the discount

hashingFactory - the hashing factory to use for storing states

vInit - the state value initialization function to use.

backupOperator - the backup operator that defines the solution concept being solved

maxDelta - the threshold that causes VI to terminate when the max Q-value change is less than it

maxIterations - the maximum number of iterations allowed

Method Detail

planFromState

public void planFromState(State s)

Description copied from class: MAValueFunctionPlanner

Calling this method causes planning to be performed from State s.

Specified by:

planFromState in class MAValueFunctionPlanner

Parameters:

s - the state from which planning is to be performed.

runVI

public void runVI()

Runs Value Iteration over the set of states that have been discovered. VI terminates either when the max change in Q-value is less than the threshold stored in this object's maxDelta parameter or when the number of iterations exceeds this object's maxIterations parameter.

If performStateReachabilityFrom(State) has not yet been called, then the state set will be empty and a runtime exception will be thrown.

performStateReachabilityFrom

public boolean performStateReachabilityFrom(State s)

Finds and stores all states that are reachable from input state s.

Parameters:

s - the state from which all reachable states will be indexed

Returns:

true if input s was not previously indexed resulting in new states being found; false if s was already previously indexed resulting in no change in the discovered state set.