public static void InitializeFromProbsArray(ref CompressedPolicyVector policyRef, int numMoves, int numMovesToSave, Span <ProbEntry> probs)
{
QuickSort(probs, 0, numMoves - 1);
// Compute max probability so we can then
// avoid overflow during exponentation by subtracting off
float max = 0.0f;
for (int j = 0; j < numMovesToSave; j++)
{
if (probs[j].P > max)
{
max = probs[j].P;
}
}
Span <float> probsA = stackalloc float[numMoves];
Span <int> indicesA = stackalloc int[numMoves];
for (int j = 0; j < numMovesToSave; j++)
{
probsA[j] = MathF.Exp(probs[j].P - max);
indicesA[j] = probs[j].Index;
}
CompressedPolicyVector.Initialize(ref policyRef, indicesA.Slice(0, numMovesToSave), probsA.Slice(0, numMovesToSave));
}
/// <summary>
/// Returns the CompressedPolicyVector which is the linear combination
/// of a set of other raw policy vectors (using a specified set of weights).
/// </summary>
/// <param name="policies"></param>
/// <param name="weights"></param>
/// <returns></returns>
public static CompressedPolicyVector LinearlyCombined(CompressedPolicyVector[] policies, float[] weights)
{
Span <float> policyAverages = stackalloc float[EncodedPolicyVector.POLICY_VECTOR_LENGTH];
// Compute average policy result for this position
for (int i = 0; i < policies.Length; i++)
{
CompressedPolicyVector policy = policies[i];
foreach ((EncodedMove move, float probability)moveInfo in policy.ProbabilitySummary())
{
if (moveInfo.move.RawValue == SPECIAL_VALUE_RANDOM_NARROW ||
moveInfo.move.RawValue == SPECIAL_VALUE_RANDOM_WIDE)
{
throw new NotImplementedException("Method LinearlyCombined probably not yet supported with random evaluations.");
}
float thisContribution = weights[i] * moveInfo.probability;
policyAverages[moveInfo.move.IndexNeuralNet] += thisContribution;
}
}
CompressedPolicyVector policyRet = default;
Initialize(ref policyRet, policyAverages, false);
return(policyRet);
}
/// <summary>
/// Initializes values (bypassing readonly) with specified set of move indices and probabilities.
/// </summary>
/// <param name="policy"></param>
/// <param name="indices"></param>
/// <param name="probs"></param>
public static void Initialize(ref CompressedPolicyVector policy, Span <ushort> indices, Span <ushort> probsEncoded)
{
if (indices.Length != probsEncoded.Length)
{
throw new ArgumentException("Length of indicies and probabilities must be same");
}
float lastProb = float.MaxValue;
fixed(ushort *moveIndices = &policy.MoveIndex_0)
{
fixed(ushort *moveProbabilitiesEncoded = &policy.MoveProbEncoded_0)
{
for (int i = 0; i < indices.Length && i < NUM_MOVE_SLOTS; i++)
{
if (indices[i] == SPECIAL_VALUE_SENTINEL_TERMINATOR)
{
break;
}
moveIndices[i] = indices[i];
moveProbabilitiesEncoded[i] = probsEncoded[i];
Debug.Assert(DecodedProbability(probsEncoded[i]) <= lastProb);
lastProb = probsEncoded[i];
}
}
}
}
/// <summary>
/// A special encoding is used to indicate if the policy is desired to be initialized
/// randomly (for testing purposes).
///
/// The actual probabilities cannot be computed here since we don't yet know the move list,
/// therefore weput a special value in the array to indicate that this should be expanded in subsequent processing.
/// </summary>
/// <param name="wide"></param>
public static void InitializeAsRandom(ref CompressedPolicyVector policy, bool wide)
{
fixed(ushort *moveIndices = &policy.MoveIndex_0)
{
// Only need to set first move index
moveIndices[0] = wide ? SPECIAL_VALUE_RANDOM_WIDE : SPECIAL_VALUE_RANDOM_NARROW;
}
}
/// <summary>
/// Compares with another CompressedPolicyVector and returns the
/// magnitude of the largest abolute difference in policy across all moves.
/// </summary>
/// <param name="other"></param>
/// <returns></returns>
public float MaxProbDifferenceWith(CompressedPolicyVector other)
{
float[] decoded = DecodedNoValidate;
float[] otherDecoded = other.DecodedNoValidate;
float max = 0;
for (int i = 0; i < EncodedPolicyVector.POLICY_VECTOR_LENGTH; i++)
{
float diff = Math.Abs(decoded[i] - otherDecoded[i]);
if (diff > max)
{
max = diff;
}
}
return(max);
}
/// <summary>
/// Static method to initialize a CompressedPolicyVector from
/// a specified array of expanded policy probabilities.
/// </summary>
/// <param name="policy"></param>
/// <param name="probabilities"></param>
/// <param name="alreadySorted"></param>
public static void Initialize(ref CompressedPolicyVector policy, float *probabilities, bool alreadySorted)
{
float probabilityAcc = 0.0f;
int numSlotsUsed = 0;
fixed(ushort *moveIndices = &policy.MoveIndex_0)
{
fixed(ushort *moveProbabilitiesEncoded = &policy.MoveProbEncoded_0)
{
// Move all the probabilities into our array
for (int i = 0; i < EncodedPolicyVector.POLICY_VECTOR_LENGTH; i++)
{
float thisProb = probabilities[i];
probabilityAcc += thisProb;
if (probabilities[i] > HALF_INCREMENT)
{
ushort encodedProb = EncodedProbability(probabilities[i]);
if (numSlotsUsed < NUM_MOVE_SLOTS)
{
moveIndices[numSlotsUsed] = (ushort)i;
moveProbabilitiesEncoded[numSlotsUsed] = encodedProb;
//Console.WriteLine("direct " + i + " " + probabilities[i]);
numSlotsUsed++;
}
else
{
// Find smallest index/value
int smallestIndex = -1;
ushort smallestValue = ushort.MaxValue;
for (int si = 0; si < NUM_MOVE_SLOTS; si++)
{
if (moveProbabilitiesEncoded[si] < smallestValue)
{
smallestIndex = si;
smallestValue = moveProbabilitiesEncoded[si];
}
}
ushort encodedSmallest = EncodedProbability(probabilities[smallestIndex]);
if (moveProbabilitiesEncoded[smallestIndex] < encodedProb)
{
moveIndices[smallestIndex] = (ushort)i;
moveProbabilitiesEncoded[smallestIndex] = encodedProb;
}
else
{
// just drop it (lost)
// Console.WriteLine(" drop " + encodedProb + " --> " + thisProb);
}
}
}
// Add terminator if not full
if (numSlotsUsed < NUM_MOVE_SLOTS)
{
moveIndices[numSlotsUsed] = SPECIAL_VALUE_SENTINEL_TERMINATOR;
}
}
if (probabilityAcc < 0.995 || probabilityAcc > 1.005)
{
throw new Exception($"Internal error: NN probabilities sum to { probabilityAcc}");
}
}
}
if (!alreadySorted)
{
policy.Sort(numSlotsUsed);
}
}
/// <summary>
/// Initializes values (bypassing readonly) with specified set of move indices and probabilities.
/// </summary>
/// <param name="policy"></param>
/// <param name="probabilities"></param>
/// <param name="alreadySorted"></param>
public static void Initialize(ref CompressedPolicyVector policy, Span <float> probabilities, bool alreadySorted)
{
Initialize(ref policy, Fixed(probabilities), alreadySorted);
}
/// <summary>
/// Initializes values (bypassing readonly) with specified set of move indices and probabilities.
/// </summary>
/// <param name="policy"></param>
/// <param name="indices"></param>
/// <param name="probs"></param>
internal static void Initialize(ref CompressedPolicyVector policy,
Span <int> indices, Span <float> probs, bool alreadySorted = true)
{
// TODO: the Span<int> can actually be shortend to Span<short>
if (indices.Length != probs.Length)
{
throw new ArgumentException("Length of indicies and probabilities must be same");
}
float probabilityAcc = 0.0f;
float priorProb = float.MaxValue; // only used in debug mode for verifying in order
int numMovesUsed = 0;
fixed(ushort *moveIndices = &policy.MoveIndex_0)
{
fixed(ushort *moveProbabilitiesEncoded = &policy.MoveProbEncoded_0)
{
// Move all the probabilities into our array
for (int i = 0; i < indices.Length && i < NUM_MOVE_SLOTS; i++)
{
// Get this probability and make sure is in expected sorted order
float thisProb = probs[i];
Debug.Assert(!alreadySorted || thisProb <= priorProb);
// Save index
int moveIndex = indices[i];
moveIndices[i] = (ushort)moveIndex;
// Save compressed probability (unless rounds to zero)
ushort encoded = EncodedProbability(thisProb);
if (encoded != 0)
{
numMovesUsed++;
moveProbabilitiesEncoded[i] = encoded;
probabilityAcc += thisProb;
}
else
{
break; // moves are sorted, so we will not see any more above MIN_VALUE
}
priorProb = thisProb;
}
if (numMovesUsed < NUM_MOVE_SLOTS)
{
// Not full. Add terminator.
moveIndices[numMovesUsed] = SPECIAL_VALUE_SENTINEL_TERMINATOR;
}
// Normalize to sum to 1.0
float adj = 1.0f / probabilityAcc;
for (int i = 0; i < numMovesUsed; i++)
{
moveProbabilitiesEncoded[i] = EncodedProbability(probs[i] * adj);
}
}
if (!alreadySorted)
{
policy.Sort(numMovesUsed);
}
}
}
