Add files via upload

New version: enhanced certainty propagation, policy compression, tree balancing, easy early visits and more

Author: Videodr0me

src/main.cc

@@ -25,7 +25,7 @@
 int main(int argc, const char** argv) {
   std::cerr << "       _" << std::endl;
   std::cerr << "|   _ | |" << std::endl;
-  std::cerr << "|_ |_ |_| built " << __DATE__ << std::endl;
+  std::cerr << "|_ |_ |_| MCTS EXPERIMENTAL " << __DATE__ << std::endl;
   using namespace lczero;
   CommandLine::Init(argc, argv);
   CommandLine::RegisterMode("uci", "(default) Act as UCI engine");

src/mcts/node.cc

@@ -206,8 +206,11 @@ void Node::ResetStats() {
   n_ = 0;
   v_ = 0.0;
   q_ = 0.0;
-  w_ = 0.0;
+  w_ = 0.0; // no longer needed
   p_ = 0.0;
+  m_ = 0.0;
+  b_ = 0.0;
+  avg_child_branches_ = 0.0;
   max_depth_ = 0;
   full_depth_ = 0;
   is_terminal_ = false;
@@ -219,7 +222,7 @@ std::string Node::DebugString() const {
   oss << "Move: " << move_.as_string() << " Term:" << is_terminal_
       << " This:" << this << " Parent:" << parent_ << " child:" << child_
       << " sibling:" << sibling_ << " P:" << p_ << " Q:" << q_ << " W:" << w_
-      << " N:" << n_ << " N_:" << n_in_flight_;
+      << " M:" << m_<< " N:" << n_ << " N_:" << n_in_flight_;
   return oss.str();
 }
 
@@ -243,6 +246,7 @@ void Node::MakeCertain(float q) {
 	is_certain_ = true;	
 	//n_ = UINT32_MAX;
 }
+// mainly used for root
 void Node::UnCertain() {
 	is_certain_ = false;
 }
@@ -255,21 +259,28 @@ bool Node::TryStartScoreUpdate() {
 void Node::CancelScoreUpdate() { --n_in_flight_; }
 
 void Node::FinalizeScoreUpdate(float v, float kBackpropagate, int kAutoextend) {
+	float q_new = 0;
 	if (!is_certain_)
 	{
-		// gamma = 1.00 corresponds to base leela q update
-		// gamma = 0.75 is my first guess and yields +25 Elo
-		// lots of room for tuning, optimum probably 0.8-09
-		q_ += (v - q_) / (std::powf(n_, kBackpropagate) + 1);
-		//  normal leela update in q + (v-q) format
-	    //	q_ += (v - q_) / ((float)n_ + 1.0f);
-
+		// Update Q:
+		// Alternatives:
+		// q_ += (v - q_) / (std::powf(n_, kBackpropagate) + 1);
+		// q_ += (v - q_) / ((n_ * kBackpropagate) + 1);
+		q_new = q_ + (v - q_) / ((float)n_ + 1.0f); //Standard update
+
+		// Update M:
+		// Sum of Squared Difference for Variance Computation
+		// this is a numerically stable online method
+		m_ += (v - q_)*(v - q_new);
+	
+	    q_ = q_new;
 		// all nodes have v_ set to NN compute or to certain value 
 		// except nodes with only one legal move (auto-extend=1)
 		// these must inherit from child
 		if (HasOnlyOneChild() && (kAutoextend == 1)) {
 			q_ = -(child_->q_);
 			v_ = -(child_->v_);
+			m_ = child_->m_;
 		} 
 
 	  // Increment N.

src/mcts/node.h

@@ -55,7 +55,6 @@ class Node {
 	Node* GetParent() const { return parent_; }
 
 	// Gets first child
-
 	Node* GetFirstChild() const { return child_; }
 
 	// Returns whether a node has children.
@@ -80,26 +79,40 @@ class Node {
 	// Returns Q if number of visits is more than 0,
 	float GetQ(float default_q) const { return n_ ? q_ : default_q; }
 	// Returns U / (Puct * N[parent])
-	float GetU() const { return p_/ (1 + n_ + n_in_flight_); }
+	float GetU() const { return p_ / (1 + n_ + n_in_flight_); }
 	// Returns value of Value Head returned from the neural net.
 	float GetV() const { return v_; }
+	// Returns the avg. of children branches (not expanded grandchildren)
+	float GetCB() const { return avg_child_branches_; }
 	// Returns value of Move probabilityreturned from the neural net.
 	// (but can be changed by adding Dirichlet noise).
 	float GetP() const { return p_; }
+	// returns branches of this node (number of childs)
+	float GetB() const {return b_;}
+	// Returns population variance of q.
+	float GetSigma2(float default_m) const { return n_>2 ? m_/(n_-1):default_m; }
 	// Returns whether the node is known to be draw/lose/win.
 	bool IsTerminal() const { return is_terminal_; }
 	// Returns whether the node is known to have a certain score
 	bool IsCertain() const { return is_certain_; }
-
-
-  uint16_t GetFullDepth() const { return full_depth_; }
-  uint16_t GetMaxDepth() const { return max_depth_; }
+    uint16_t GetFullDepth() const { return full_depth_; }
+    uint16_t GetMaxDepth() const { return max_depth_; }
   // makes node uncertain again (used to make root uncertain when search is initialized
   void UnCertain();
+  // Sets node avg of all childrens branches
+  void SetCB(float val) { avg_child_branches_ = val; }
   // Sets node own value (from neural net or win/draw/lose adjudication).
   void SetV(float val) { v_ = val; }
   // Sets move probability.
   void SetP(float val) { p_ = val; }
+  // Sets Q
+  void SetQ(float val) { q_ = val; }
+  // Sets branches (number of childs)
+  void SetB(float val) { b_ = val; }
+  // Sets n_ for terminal nodes that are 
+  // found when creating children in 
+  // expand node
+  void SetN1() { n_ = 1; }
   // Makes the node terminal and sets it's score.
   void MakeTerminal(GameResult result);
   // Makes the node certain and sets it's score
@@ -117,9 +130,11 @@ class Node {
   // Updates:
   // * N (+=1)
   // * N-in-flight (-=1)
-  // * W (+= v)
-  // * Q (=w/n)
-  // Backpropagete and Autoextend modes are currently passed as parameters
+  // * W (+= v)  obsolete
+  // * Q (+= q + (v - q) (n_+1))
+  // * M Sum of Squares of Differences
+  // kBackpropagate (not used currently) and Autoextend modes are 
+  // currently passed as parameters
   // will either be removed if changes become permanent, or replaced
   // by a weight parameter.
   void FinalizeScoreUpdate(float v, float kBackpropagate, int kAutoextend);
@@ -171,6 +186,11 @@ class Node {
   // Probabality that this move will be made. From policy head of the neural
   // network.
   float p_;
+  // Sum of Squares of Differences from current mean
+  float m_;
+  // branch data for tree shaping
+  float b_;
+  float avg_child_branches_;
   // How many completed visits this node had.
   uint32_t n_;
   // (aka virtual loss). How many threads currently process this node (started