public override int CompareTo(IBinaryHeapItem other)
{
this.h += this.targetDeltaF;
WorldStateForPartialExpansion otherNode = (WorldStateForPartialExpansion)other;
otherNode.h += otherNode.targetDeltaF;
int res = base.CompareTo(other);
this.h -= this.targetDeltaF;
otherNode.h -= otherNode.targetDeltaF;
return(res);
}
/// <summary>
/// Copy constructor
/// </summary>
/// <param name="cpy"></param>
public WorldStateForPartialExpansion(WorldStateForPartialExpansion cpy)
: base(cpy)
{
alreadyExpanded = false; // Creating a new unexpanded node from cpy
// For intermediate nodes created during expansion (fully expanded nodes have these fields recalculated when they're expanded)
remainingDeltaF = cpy.remainingDeltaF;
singleAgentDeltaFs = cpy.singleAgentDeltaFs; // For the UpdateRemainingDeltaF call on temporary nodes.
// Notice that after an agent is moved its row won't be up-to-date.
fLookup = cpy.fLookup; // For the hasChildrenForCurrentDeltaF call on temporary nodes.
// Notice that after an agent is moved, all rows up to and including the one of the agent that moved won't be up-to-date.
maxDeltaF = cpy.maxDeltaF; // Not necessarily achievable after some of the agents moved.
// The above is OK because we won't be using data for agents that already moved.
}
public override IBinaryHeapItem Remove()
{
WorldState node;
if (this.lastF != -1)
{
this.numExpands++;
double expandFinishTime = this.runner.ElapsedMilliseconds();
this.sumExpandTimes += expandFinishTime - this.expandStartTime;
}
if (base.Count == 1) // Can happen more than once with the same node if partial expansion pushes it back into the open list
{
node = (WorldState)base.Remove();
goto finish;
}
// There are alternatives to the lowest cost node in the open list, try to postpone expansion of it:
float branchingFactor = ((ClassicAStar)this.user).GetEffectiveBranchingFactor(); // We know the solver is an A* variant.
const double binaryHeapTau = 0.073359375; // microseconds. From empirical experiments with this infra on my computer.
double logN = Math.Log(this.heap.Count, 2); // Removals from and insertions to the queue cost practically zero.
double t0 = binaryHeapTau * logN; // TODO: Measure this directly?
double overhead = 0.023 * ((WorldState)this.Peek()).allAgentsState.Length; // in milliseconds. Empirical lowest estimate. The cost of a zero-timeout CBSH run wasn't simply linear with the number of agents for some reason.
while (true)
{
WorldState next;
node = (WorldState)base.Remove();
if (node.GoalTest() == true || // Can't improve the h of the goal
this.runner.ElapsedMilliseconds() > Constants.MAX_TIME) // No time to continue improving H.
{
if (node.g + node.h < this.lastF) // This can happen if the last removed node had many runs of the expensive heuristic, which this node didn't yet have.
{
int newH = this.lastF - node.g;
node.hBonus += newH - node.h;
node.h = newH; // Just so we don't throw an inconsistency exception
}
break;
}
if (node.hBonus > 0) // Improving the h will be more difficult than usual
{
this.skips++; // TODO: Separate statistic?
break;
}
int selectedCapExponent = -1;
if (runOracle)
{
this.runner.StartOracle();
next = (WorldState)base.Peek();
int targetH = next.g + next.h + 1 - node.g;
Stopwatch watch = Stopwatch.StartNew();
WorldStateForPartialExpansion nodeCopy = new WorldStateForPartialExpansion((WorldStateForPartialExpansion)node); // So the solution won't leak to the node when we try to solve it.
double expensiveCallStartTime = watch.Elapsed.TotalMilliseconds;
// Using a separate statistic for the oracle to keep the stats clean
int expensiveEstimate = (int)((LazyHeuristic)this.runner.heuristics[this.runner.heuristics.Count - 1]).h(
nodeCopy, targetH, -1, int.MaxValue, false);
double expensiveCallTotalTime = watch.Elapsed.TotalMilliseconds - expensiveCallStartTime;
int lowestCapThatWouldHaveWorked = (int)Math.Ceiling(Math.Log(expensiveCallTotalTime * 1000, 2));
//Console.WriteLine("Lowest cap that would have worked:{0}", (1<<lowestCapThatWouldHaveWorked)/1000.0);
for (int j = 0; (j < lowestCapThatWouldHaveWorked) && (j < DynamicRationalLazyOpenList.NUM_CAPS); ++j)
{
this.capData[j, DynamicRationalLazyOpenList.FAILURE_IND] += 1;
this.capData[j, DynamicRationalLazyOpenList.PH_IND] = 0; // Correct for this specific node.
// All expanded nodes either have the oracle run on them or they don't, so this value won't leak to other nodes.
}
for (int j = lowestCapThatWouldHaveWorked; j < DynamicRationalLazyOpenList.NUM_CAPS; ++j)
{
this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] = +1;
this.capData[j, DynamicRationalLazyOpenList.PH_IND] = 1;
}
this.runner.StopOracle();
}
// DRLA* calculation (derived from Tolpin et al. RLA* formula):
double tExpand = (this.sumExpandTimes / this.numExpands) * 1000; // in microseconds
int nextCapExponentAboveTExpand;
if (tExpand != 0)
{
nextCapExponentAboveTExpand = (int)Math.Ceiling(Math.Log(tExpand, 2));
nextCapExponentAboveTExpand = Math.Min(nextCapExponentAboveTExpand, DynamicRationalLazyOpenList.NUM_CAPS - 1);
}
else
{
nextCapExponentAboveTExpand = 0;
}
// Force the heuristic to be tried if it's never been tried before at a cap that's one above Texpand.
// This is needed because we initially believe all caps are useless, so the caps that are too short won't be tried
// and fail. This has the added benefit that after running, we'll have an informed belief for all lower caps and possibly
// all higher caps too.
if (this.capData[nextCapExponentAboveTExpand, DynamicRationalLazyOpenList.SUCCESS_IND] +
this.capData[nextCapExponentAboveTExpand, DynamicRationalLazyOpenList.FAILURE_IND] == 0) // This cap was never tried before, and it might be good.
{
selectedCapExponent = nextCapExponentAboveTExpand;
}
else
{
double[] expectedRegret = new double[DynamicRationalLazyOpenList.NUM_CAPS]; // in microseconds
double lowestRegret = double.MaxValue;
//double tempStartTime = ((DyanamicLazyCbsh)this.expensive).runner.ElapsedMilliseconds();
// Computing the expected regret from choosing each cap, and finding the cap that gives the lowest expected regret
for (int i = 0; i < DynamicRationalLazyOpenList.NUM_CAPS; ++i)
{
expectedRegret[i] = 0;
// Going over the probabilities that the caps would be helpful
for (int j = 0; j <= i; ++j)
{
double heuristicOnlyHelpfulAfter = 1 << j; // microseconds
double probabilityForThat;
if (j != 0)
{
probabilityForThat = this.capData[j, DynamicRationalLazyOpenList.PH_IND] - this.capData[j - 1, DynamicRationalLazyOpenList.PH_IND]; // The PH_IND points to the probability that the given cap would succeed at all, including the probability that it would've worked for lower caps.
}
else
{
probabilityForThat = this.capData[j, DynamicRationalLazyOpenList.PH_IND];
}
expectedRegret[i] += (0.75 * heuristicOnlyHelpfulAfter + t0 - Math.Min(0.75 * heuristicOnlyHelpfulAfter + t0, tExpand)) * probabilityForThat; // The cap is an upper bound on t2. A closer estimate is 0.75*cap (since 0.5*cap is the previous cap).
}
double cap = 1 << i;
for (int j = i + 1; j < DynamicRationalLazyOpenList.NUM_CAPS; ++j)
{
double heuristicOnlyHelpfulAfter = 1 << j; // microseconds
double probabilityForThat = this.capData[j, DynamicRationalLazyOpenList.PH_IND] - this.capData[j - 1, DynamicRationalLazyOpenList.PH_IND];
expectedRegret[i] += (cap + tExpand - Math.Min(0.75 * heuristicOnlyHelpfulAfter + t0, tExpand)) * probabilityForThat;
}
// The case where even the highest cap fails:
int last = DynamicRationalLazyOpenList.NUM_CAPS - 1;
expectedRegret[i] += cap * (1 - this.capData[last, DynamicRationalLazyOpenList.PH_IND]);
if (expectedRegret[i] < lowestRegret)
{
lowestRegret = expectedRegret[i];
selectedCapExponent = i;
}
}
}
double millisCap = ((double)(1 << selectedCapExponent)) / 1000;
//double tempFinishTime = ((DyanamicLazyCbsh)this.expensive).runner.ElapsedMilliseconds();
//Console.WriteLine("calculations time:{0}", tempFinishTime - tempStartTime);
//Console.WriteLine("millisCap:{0}", millisCap);
//Console.WriteLine("Texpand:{0}={1}/{2}", tExpand, this.sumExpandTimes * 1000, this.numExpands);
//Console.WriteLine("T0:{0}", t0);
//for (int i=0 ; i<DynamicRationalLazyOpenList.NUM_CAPS ; ++i)
//Console.WriteLine("Ph for cap {0}:{1}={2}/{3}", ((double)(1 << i)) / 1000, this.capData[i, 2], this.capData[i, 0], this.capData[i, 0] + this.capData[i, 1]);
//for (int i = 0; i < DynamicRationalLazyOpenList.NUM_CAPS; ++i)
//Console.WriteLine("Expected regret for cap {0}:{1}", ((double)(1 << i)) / 1000, expectedRegret[i]/1000.0);
if (node.g + node.h < lastF) // Must improve the heuristic estimate to be consistent
{
millisCap = Double.MaxValue;
}
bool success = false;
if (millisCap > overhead) // Worth running the expensive heuristic
{
next = (WorldState)base.Peek();
int targetH = next.g + next.h + 1 - node.g;
double expensiveCallStartTime = this.runner.ElapsedMilliseconds();
int expensiveEstimate = (int)this.expensive.h(node, targetH, -1, (int)(expensiveCallStartTime + millisCap), false);
double expensiveCallTotalTime = this.runner.ElapsedMilliseconds() - expensiveCallStartTime;
bool nodeSolved = node.GoalTest();
if (expensiveEstimate > node.h) // Node may have inherited a better estimate from its parent so this check is necessary for failures
{
node.hBonus += expensiveEstimate - node.h;
node.h = expensiveEstimate; // If this wasn't a success, the assignment here serves to force the next heuristic call
// to search deeper, since we're always consistent.
}
this.capData[selectedCapExponent, DynamicRationalLazyOpenList.USES_IND]++;
success = node.CompareTo(next) == 1; // Node is not the smallest F anymore - re-insert it into the open list
// Update capData:
if (success || nodeSolved)
{
int lowestCapThatWouldHaveWorked = (int)Math.Ceiling(Math.Log(expensiveCallTotalTime * 1000, 2));
//Console.WriteLine("Lowest cap that would have worked:{0}", (1<<lowestCapThatWouldHaveWorked)/1000.0);
for (int j = 0; (j < lowestCapThatWouldHaveWorked) && (j < DynamicRationalLazyOpenList.NUM_CAPS); ++j)
{
this.capData[j, DynamicRationalLazyOpenList.FAILURE_IND] += 1;
this.capData[j, DynamicRationalLazyOpenList.PH_IND] = /*(1 - DynamicRationalLazyOpenList.MOVING_AVERAGE_FACTOR) * this.capData[j, DynamicRationalLazyOpenList.PH_IND] +
* DynamicRationalLazyOpenList.MOVING_AVERAGE_FACTOR * */this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] / (this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] + this.capData[j, DynamicRationalLazyOpenList.FAILURE_IND]);
}
for (int j = lowestCapThatWouldHaveWorked; j < DynamicRationalLazyOpenList.NUM_CAPS; ++j)
{
this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] += 1;
this.capData[j, DynamicRationalLazyOpenList.PH_IND] = /*(1 - DynamicRationalLazyOpenList.MOVING_AVERAGE_FACTOR) * this.capData[j, DynamicRationalLazyOpenList.PH_IND] +
* DynamicRationalLazyOpenList.MOVING_AVERAGE_FACTOR * */this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] / (this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] + this.capData[j, DynamicRationalLazyOpenList.FAILURE_IND]);
}
}
else
{
int effectiveCap = (int)Math.Floor(Math.Log(expensiveCallTotalTime * 1000, 2)); // Heuristic may have gone over its timeout
for (int j = 0; (j <= effectiveCap) && (j < DynamicRationalLazyOpenList.NUM_CAPS); ++j)
{
this.capData[j, DynamicRationalLazyOpenList.FAILURE_IND] += 1;
this.capData[j, DynamicRationalLazyOpenList.PH_IND] = /*(1 - DynamicRationalLazyOpenList.MOVING_AVERAGE_FACTOR) * this.capData[j, DynamicRationalLazyOpenList.PH_IND] +
* DynamicRationalLazyOpenList.MOVING_AVERAGE_FACTOR * */this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] / (this.capData[j, DynamicRationalLazyOpenList.SUCCESS_IND] + this.capData[j, DynamicRationalLazyOpenList.FAILURE_IND]); // This is only necessary for the first runs so the Ph won't jitter too much.
}
// No info on whether a larger cap would have worked.
}
}
else
{
this.skips++;
}
if (success || node.g + node.h < lastF) // Never be inconsistent - don't return nodes with lower F than before. Try searching the node again.
{
this.Add(node);
}
else
{
// Node is still less than or equal to all items in the open list (and its h would look consistent to A*).
// This can be because of many reasons:
// - The search ended because of a timeout
// - The search ended because a goal was found (good!)
// - The search ended because the CBS search was too costly
break;
}
}
finish:
this.lastF = node.g + node.h;
this.expandStartTime = this.runner.ElapsedMilliseconds();
return(node);
}
override protected WorldState CreateSearchRoot(int minDepth = -1)
{
WorldStateForPartialExpansion root = new WorldStateForPartialExpansion(this.instance.m_vAgents, minDepth);
return(root);
}
