s14h
/
prism-accumulation
1
0
Fork 0
Code Releases Activity
Browse Source

Enable simulation of POMDPs via SimulatorEngine.

accumulation-v4.7
Dave Parker 5 years ago
parent
e5ea762d80
commit
d78b627bf4
8 changed files with 100 additions and 20 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 6
      prism/src/cex/CexPathAsBDDs.java
  2. 6
      prism/src/cex/CexPathStates.java
  3. 15
      prism/src/simulator/Path.java
  4. 27
      prism/src/simulator/PathFull.java
  5. 6
      prism/src/simulator/PathFullInfo.java
  6. 12
      prism/src/simulator/PathFullPrefix.java
  7. 25
      prism/src/simulator/PathOnTheFly.java
  8. 23
      prism/src/simulator/SimulatorEngine.java

6
prism/src/cex/CexPathAsBDDs.java
View File

@ -85,6 +85,12 @@ public class CexPathAsBDDs implements PathFullInfo
return model.convertBddToState(states.get(step));
}
@Override
public State getObservation(int step)
{
return null;
}
@Override
public double getStateReward(int step, int rsi)
{

6
prism/src/cex/CexPathStates.java
View File

@ -79,6 +79,12 @@ public class CexPathStates implements PathFullInfo
return states.get(step);
}
@Override
public State getObservation(int step)
{
return null;
}
@Override
public double getStateReward(int step, int rsi)
{

15
prism/src/simulator/Path.java
View File

@ -37,22 +37,22 @@ public abstract class Path
// MUTATORS
/**
* Initialise the path with an initial state and rewards.
* Note: State object and array will be copied, not stored directly.
* Initialise the path with an initial state, observation and rewards.
* Note: State objects and array will be copied, not stored directly.
*/
public abstract void initialise(State initialState, double[] initialStateRewards);
public abstract void initialise(State initialState, State initialObs, double[] initialStateRewards);
/**
* Add a step to the path.
* Note: State object and arrays will be copied, not stored directly.
*/
public abstract void addStep(int choice, Object action, String actionString, double probability, double[] transRewards, State newState, double[] newStateRewards, ModelGenerator modelGen);
public abstract void addStep(int choice, Object action, String actionString, double probability, double[] transRewards, State newState, State newObs, double[] newStateRewards, ModelGenerator modelGen);
/**
* Add a timed step to the path.
* Note: State object and arrays will be copied, not stored directly.
*/
public abstract void addStep(double time, int choice, Object action, String actionString, double probability, double[] transRewards, State newState, double[] newStateRewards, ModelGenerator modelGen);
public abstract void addStep(double time, int choice, Object action, String actionString, double probability, double[] transRewards, State newState, State newObs, double[] newStateRewards, ModelGenerator modelGen);
// ACCESSORS
@ -76,6 +76,11 @@ public abstract class Path
*/
public abstract State getCurrentState();
/**
* Get the observation for the current state, i.e. for the current final state of the path.
*/
public abstract State getCurrentObservation();
/**
* Get the action taken in the previous step.
*/

27
prism/src/simulator/PathFull.java
View File

@ -89,7 +89,7 @@ public class PathFull extends Path implements PathFullInfo
// MUTATORS (for Path)
@Override
public void initialise(State initialState, double[] initialStateRewards)
public void initialise(State initialState, State initialObs, double[] initialStateRewards)
{
clear();
// Add new step item to the path
@ -97,6 +97,7 @@ public class PathFull extends Path implements PathFullInfo
steps.add(step);
// Add (copies of) initial state and state rewards to new step
step.state = new State(initialState);
step.obs = initialObs == null ? null : new State(initialObs);
step.stateRewards = initialStateRewards.clone();
// Set cumulative time/reward (up until entering this state)
step.timeCumul = 0.0;
@ -108,15 +109,15 @@ public class PathFull extends Path implements PathFullInfo
}
@Override
public void addStep(int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState, double[] newStateRewards,
public void addStep(int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState, State newObs, double[] newStateRewards,
ModelGenerator modelGen)
{
addStep(1.0, choice, action, actionString, probability, transitionRewards, newState, newStateRewards, modelGen);
addStep(1.0, choice, action, actionString, probability, transitionRewards, newState, newObs, newStateRewards, modelGen);
}
@Override
public void addStep(double time, int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState,
double[] newStateRewards, ModelGenerator modelGen)
State newObs, double[] newStateRewards, ModelGenerator modelGen)
{
Step stepOld, stepNew;
// Add info to last existing step
@ -130,8 +131,9 @@ public class PathFull extends Path implements PathFullInfo
// Add new step item to the path
stepNew = new Step();
steps.add(stepNew);
// Add (copies of) new state and state rewards to new step
// Add (copies of) new state, observation and state rewards to new step
stepNew.state = new State(newState);
stepNew.obs = newObs == null ? null : new State(newObs);
stepNew.stateRewards = newStateRewards.clone();
// Set cumulative time/rewards (up until entering this state)
stepNew.timeCumul = stepOld.timeCumul + time;
@ -241,6 +243,12 @@ public class PathFull extends Path implements PathFullInfo
return steps.get(steps.size() - 1).state;
}
@Override
public State getCurrentObservation()
{
return steps.get(steps.size() - 1).obs;
}
@Override
public Object getPreviousAction()
{
@ -339,6 +347,12 @@ public class PathFull extends Path implements PathFullInfo
return steps.get(step).state;
}
@Override
public State getObservation(int step)
{
return steps.get(step).obs;
}
@Override
public double getStateReward(int step, int rsi)
{
@ -545,6 +559,7 @@ public class PathFull extends Path implements PathFullInfo
{
// Set (unknown) defaults and initialise arrays
state = null;
obs = null;
stateRewards = new double[numRewardStructs];
timeCumul = 0.0;
rewardsCumul = new double[numRewardStructs];
@ -558,6 +573,8 @@ public class PathFull extends Path implements PathFullInfo
// Current state (before transition)
public State state;
// Observation for current state
public State obs;
// State rewards for current state
public double stateRewards[];
// Cumulative time spent up until entering this state

6
prism/src/simulator/PathFullInfo.java
View File

@ -44,6 +44,12 @@ public interface PathFullInfo
*/
public abstract State getState(int step);
/**
* Get the observation at a given step of the path.
* @param step Step index (0 = initial state/step of path)
*/
public abstract State getObservation(int step);
/**
* Get a state reward for the state at a given step of the path.
* If no reward info is stored ({@link #hasRewardInfo()} is false), returns 0.0.

12
prism/src/simulator/PathFullPrefix.java
View File

@ -52,19 +52,19 @@ public class PathFullPrefix extends Path
// MUTATORS (for Path)
@Override
public void initialise(State initialState, double[] initialStateRewards)
public void initialise(State initialState, State initialObs, double[] initialStateRewards)
{
// Do nothing (we are not allowed to modify the underlying PathFull)
}
@Override
public void addStep(int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState, double[] newStateRewards, ModelGenerator modelGen)
public void addStep(int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState, State newObs, double[] newStateRewards, ModelGenerator modelGen)
{
// Do nothing (we are not allowed to modify the underlying PathFull)
}
@Override
public void addStep(double time, int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState, double[] newStateRewards, ModelGenerator modelGen)
public void addStep(double time, int choice, Object action, String actionString, double probability, double[] transitionRewards, State newState, State newObs, double[] newStateRewards, ModelGenerator modelGen)
{
// Do nothing (we are not allowed to modify the underlying PathFull)
}
@ -102,6 +102,12 @@ public class PathFullPrefix extends Path
return pathFull.getState(prefixLength);
}
@Override
public State getCurrentObservation()
{
return pathFull.getObservation(prefixLength);
}
@Override
public Object getPreviousAction()
{

25
prism/src/simulator/PathOnTheFly.java
View File

@ -48,6 +48,7 @@ public class PathOnTheFly extends Path
protected long size;
protected State previousState;
protected State currentState;
protected State currentObs;
protected Object previousAction;
protected String previousActionString;
protected double previousProbability;
@ -73,6 +74,10 @@ public class PathOnTheFly extends Path
// Create State objects for current/previous state
previousState = new State(modelInfo.getNumVars());
currentState = new State(modelInfo.getNumVars());
currentObs = null;
if (modelInfo.getModelType().partiallyObservable()) {
currentObs = new State(modelInfo.getNumObservables());
}
// Create arrays to store totals
totalRewards = new double[numRewardStructs];
previousStateRewards = new double[numRewardStructs];
@ -106,10 +111,13 @@ public class PathOnTheFly extends Path
// MUTATORS (for Path)
@Override
public void initialise(State initialState, double[] initialStateRewards)
public void initialise(State initialState, State initialObs, double[] initialStateRewards)
{
clear();
currentState.copy(initialState);
if (initialObs != null) {
currentObs.copy(initialObs);
}
for (int i = 0; i < numRewardStructs; i++) {
currentStateRewards[i] = initialStateRewards[i];
}
@ -118,17 +126,20 @@ public class PathOnTheFly extends Path
}
@Override
public void addStep(int choice, Object action, String actionString, double probability, double[] transRewards, State newState, double[] newStateRewards, ModelGenerator modelGen)
public void addStep(int choice, Object action, String actionString, double probability, double[] transRewards, State newState, State newObs, double[] newStateRewards, ModelGenerator modelGen)
{
addStep(1.0, choice, action, actionString, probability, transRewards, newState, newStateRewards, modelGen);
addStep(1.0, choice, action, actionString, probability, transRewards, newState, newObs, newStateRewards, modelGen);
}
@Override
public void addStep(double time, int choice, Object action, String actionString, double probability, double[] transRewards, State newState, double[] newStateRewards, ModelGenerator modelGen)
public void addStep(double time, int choice, Object action, String actionString, double probability, double[] transRewards, State newState, State newObs, double[] newStateRewards, ModelGenerator modelGen)
{
size++;
previousState.copy(currentState);
currentState.copy(newState);
if (newObs != null) {
currentObs.copy(newObs);
}
previousAction = action;
previousActionString = actionString;
previousProbability = probability;
@ -174,6 +185,12 @@ public class PathOnTheFly extends Path
return currentState;
}
@Override
public State getCurrentObservation()
{
return currentObs;
}
@Override
public Object getPreviousAction()
{

23
prism/src/simulator/SimulatorEngine.java
View File

@ -303,10 +303,12 @@ public class SimulatorEngine extends PrismComponent
throw new PrismNotSupportedException("Random choice of multiple initial states not yet supported");
}
}
// Get initial observation
State currentObs = modelGen.getObservation(currentState);
// Get initial state reward
calculateStateRewards(currentState, tmpStateRewards);
// Initialise stored path
path.initialise(currentState, tmpStateRewards);
path.initialise(currentState, currentObs, tmpStateRewards);
// Explore initial state in model generator
computeTransitionsForState(currentState);
// Reset and then update samplers for any loaded properties
@ -373,6 +375,7 @@ public class SimulatorEngine extends PrismComponent
executeTransition(ref.i, ref.offset, -1);
break;
case MDP:
case POMDP:
// Pick a random choice
i = rng.randomUnifInt(modelGen.getNumChoices());
// Pick a random transition from this choice
@ -858,10 +861,12 @@ public class SimulatorEngine extends PrismComponent
calculateTransitionRewards(path.getCurrentState(), action, tmpTransitionRewards);
// Compute next state
currentState.copy(modelGen.computeTransitionTarget(i, offset));
// Compute observation for new state
State currentObs = modelGen.getObservation(currentState);
// Compute state rewards for new state
calculateStateRewards(currentState, tmpStateRewards);
// Update path
path.addStep(index, action, actionString, p, tmpTransitionRewards, currentState, tmpStateRewards, modelGen);
path.addStep(index, action, actionString, p, tmpTransitionRewards, currentState, currentObs, tmpStateRewards, modelGen);
// Explore new state in model generator
computeTransitionsForState(currentState);
// Update samplers for any loaded properties
@ -896,10 +901,12 @@ public class SimulatorEngine extends PrismComponent
calculateTransitionRewards(path.getCurrentState(), action, tmpTransitionRewards);
// Compute next state
currentState.copy(modelGen.computeTransitionTarget(i, offset));
// Compute observation for new state
State currentObs = modelGen.getObservation(currentState);
// Compute state rewards for new state
calculateStateRewards(currentState, tmpStateRewards);
// Update path
path.addStep(time, index, action, actionString, p, tmpTransitionRewards, currentState, tmpStateRewards, modelGen);
path.addStep(time, index, action, actionString, p, tmpTransitionRewards, currentState, currentObs, tmpStateRewards, modelGen);
// Explore new state in model generator
computeTransitionsForState(currentState);
// Update samplers for any loaded properties
@ -1278,6 +1285,16 @@ public class SimulatorEngine extends PrismComponent
return ((PathFull) path).getState(step);
}
/**
* Get the observation at a given step of the path.
* (Not applicable for on-the-fly paths)
* @param step Step index (0 = initial state/step of path)
*/
public State getObservationOfPathStep(int step)
{
return ((PathFull) path).getObservation(step);
}
/**
* Get a state reward for the state at a given step of the path.
* (Not applicable for on-the-fly paths)

Write Preview
Loading…
Cancel
Save