burlap.behavior.singleagent.learnfromdemo.mlirl

Class MLIRL

java.lang.Object

burlap.behavior.singleagent.learnfromdemo.mlirl.MLIRL

public class MLIRL
extends java.lang.Object

An implementation of Maximum-likelihood Inverse Reinforcement Learning [1]. This class takes as input (from an MLIRLRequest object) a set of expert trajectories through a domain and a DifferentiableRF model, and learns the parameters of the reward function model that maximizes the likelihood of the trajectories. The reward function parameter spaces is searched using gradient ascent. Since the policy gradient it uses is non-linear, it's possible that it may get stuck in local optimas. Computing the policy gradient is done by iteratively replanning after each gradient ascent step with a QGradientPlanner instance provided in the MLIRLRequest object.

The gradient ascent will stop either after a fixed number of steps or until the change in likelihood is smaller than some threshold. If the max number of steps is set to -1, then it will continue until the change in likelihood is smaller than the threshold.

1. Babes, Monica, et al. "Apprenticeship learning about multiple intentions." Proceedings of the 28th International Conference on Machine Learning (ICML-11). 2011.

Author:

James MacGlashan.

Field Summary

Fields

Modifier and Type	Field and Description
protected int	debugCode The debug code used for printing information to the terminal.
protected double	learningRate The gradient ascent learning rate
protected double	maxLikelihoodChange The likelihood change threshold to stop gradient ascent.
protected int	maxSteps The maximum number of steps of gradient ascent.
protected MLIRLRequest	request The MLRIL request defining the IRL problem.

Constructor Summary

Constructors

Constructor and Description
MLIRL(MLIRLRequest request, double learningRate, double maxLikelihoodChange, int maxSteps) Initializes.

Method Summary

Methods

Modifier and Type	Method and Description
protected static void	addToVector(double[] sumVector, double[] deltaVector) Performs a vector addition and stores the results in sumVector
int	getDebugCode() Returns the debug code used for printing to the terminal
double	logLikelihood() Computes and returns the log-likelihood of all expert trajectories under the current reward function parameters.
FunctionGradient	logLikelihoodGradient() Computes and returns the gradient of the log-likelihood of all trajectories
double	logLikelihoodOfTrajectory(EpisodeAnalysis ea, double weight) Computes and returns the log-likelihood of the given trajectory under the current reward function parameters and weights it by the given weight.
FunctionGradient	logPolicyGrad(State s, GroundedAction ga) Computes and returns the gradient of the Boltzmann policy for the given state and action.
void	performIRL() Runs gradient ascent.
void	setDebugCode(int debugCode) Sets the debug code used for printing to the terminal
void	setRequest(MLIRLRequest request) Sets the MLIRLRequest object defining the IRL problem.
void	toggleDebugPrinting(boolean printDebug) Sets whether information during learning is printed to the terminal.

Methods inherited from class java.lang.Object

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

Field Detail

request

protected MLIRLRequest request

The MLRIL request defining the IRL problem.

learningRate

protected double learningRate

The gradient ascent learning rate

maxLikelihoodChange

protected double maxLikelihoodChange

The likelihood change threshold to stop gradient ascent.

maxSteps

protected int maxSteps

The maximum number of steps of gradient ascent. when set to -1, there is no limit and termination will be based on the maxLikelihoodChange alone.

debugCode

protected int debugCode

The debug code used for printing information to the terminal.

Constructor Detail

MLIRL

public MLIRL(MLIRLRequest request,
     double learningRate,
     double maxLikelihoodChange,
     int maxSteps)

Initializes.

Parameters:

request - the problem request definition

learningRate - the gradient ascent learning rate

maxLikelihoodChange - the likelihood change threshold that must be reached to terminate gradient ascent

maxSteps - the maximum number of gradient ascent steps allowed before termination is forced. Set to -1 to rely only on likelihood threshold.

Method Detail

setRequest

public void setRequest(MLIRLRequest request)

Sets the MLIRLRequest object defining the IRL problem.

Parameters:

request - the MLIRLRequest object defining the IRL problem.

toggleDebugPrinting

public void toggleDebugPrinting(boolean printDebug)

Sets whether information during learning is printed to the terminal. Will automatically toggle the debug printing for the underlying valueFunction as well.

Parameters:

printDebug - if true, information is printed to the terminal; if false then it is silent.

getDebugCode

public int getDebugCode()

Returns the debug code used for printing to the terminal

Returns:

the debug code used for printing to the terminal.

setDebugCode

public void setDebugCode(int debugCode)

Sets the debug code used for printing to the terminal

Parameters:

debugCode - the debug code used for printing to the terminal

performIRL

public void performIRL()

Runs gradient ascent.

logLikelihood

public double logLikelihood()

Computes and returns the log-likelihood of all expert trajectories under the current reward function parameters.

Returns:

the log-likelihood of all expert trajectories under the current reward function parameters.

logLikelihoodOfTrajectory

public double logLikelihoodOfTrajectory(EpisodeAnalysis ea,
                               double weight)

Computes and returns the log-likelihood of the given trajectory under the current reward function parameters and weights it by the given weight.

Parameters:

ea - the trajectory

weight - the weight to assign the trajectory

Returns:

the log-likelihood of the given trajectory under the current reward function parameters and weights it by the given weight.

logLikelihoodGradient

public FunctionGradient logLikelihoodGradient()

Computes and returns the gradient of the log-likelihood of all trajectories

Returns:

the gradient of the log-likelihood of all trajectories

logPolicyGrad

public FunctionGradient logPolicyGrad(State s,
                             GroundedAction ga)

Computes and returns the gradient of the Boltzmann policy for the given state and action.

Parameters:

s - the state in which the policy is queried

ga - the action for which the policy is queried.

Returns:

s the gradient of the Boltzmann policy for the given state and action.

addToVector

protected static void addToVector(double[] sumVector,
               double[] deltaVector)

Performs a vector addition and stores the results in sumVector

Parameters:

sumVector - the input vector to which the values in deltaVector will be added.

deltaVector - the vector values to add to sumVector.