Update README.md

Author: Videodr0me

README.md

@@ -1,6 +1,7 @@
 # leela-chess-experimental
 based on Leela Chess Zero https://github.com/LeelaChessZero
 
+Update 19/06/2018: V2.1, new source, new executable, changes to tree balancing, easy-early-visits
 Update 17/06/2018: New source, new executable, new parameters and new test results.
 
 I tried a number of MCTS search ideas in lc0. If you find something interesting here feel free to open an issue and discuss. This is a work in progress and purely experimental - new ideas will be added from time to time. This serves as documentation of both the good as well as the bad tries - so do not expect huge gains - but some ideas yield a measurable elo gain.  
@@ -9,7 +10,7 @@ Disclaimer: All changes are completely zero and game agnostic.
 
 ## Search Modifications
 
-### Tree Balancing
+### Tree Balancing - Work in Progress
 The upper confidence bound used in LC0's UCT flavor (and that of A0) assumes that the confidence bound of a child node is not affected by the local branching factor. However, in some games like Draughts or Chess the number of legal moves (branches) can vary greatly even in the same part of the search tree. This modification is based on the idea that we can use the number of individual branches in relation to the average number of branches to adjust the upper bound when selecting child nodes for expansion.  
 
 Initial testing with Parameters:
@@ -22,7 +23,13 @@ at 800 visits per move in 1000 games yielded these results:
 P1: +212 -122 =666 Win: 54.50% Elo: 31.35 LOS: 100.00% P1-W: +126 -49 =325 P1-B: +86 -73 =341
 ```
 But more testing is needed at higher visit searches. More results will follow.
-  
+ 
+Update: New Parameterset, these values roughly correspond to the above settings:
+--tree-balance=1.5
+--tree-scale-left=1.5
+--tree-scale-right=0.01
+
+More tests to follow
 
 ### Certainty propagation & Single-legal-move extension
 
@@ -35,6 +42,9 @@ P1: +181 -152 =667 Win: 51.45% Elo: 10.08 LOS: 94.40% P1-W: +102 -69 =328 P1-B:
 ```
 Besides the small gain of elo, this has some additonal nice properties. Leela now finds shallow mates faster and certain winning moves at root can be played regardless of visit counts, which is beneficial in time pressure situations (typically MCTS is slow to revise initial estimates).
 
+Update: 
+Improvement (v2.1) - if current best move is a certain loss change best move to a non loosing move, even if visits are lower.
+
 ### Compress low policy move probabilites
 Instead of changing softmax temperature this scheme encourages exploration of low policy priors by compressing low probabilites more than high probabilites in relation to search depths. This does well at tactics (>170/200 WAC Silvertestsuite with standard cpuct=1.2) but suffers somewhat in selfplay, even though results against different opponents (non leela) are better. Might be useful for long analysis as it restores MCTS convergence properties (under some circumstances leela would never find moves no matter how many nodes visited.). 
 ```
@@ -50,6 +60,8 @@ Tweaks the formula slightly to encourage early visits. The confidence bound is a
 ```
 Might help ever so slightly tactically - this is untested but might work well in conjunction with policy compression. Self-play might suffer, but untested against non-leela opponents. 
 
+Update: This is now float parameter with 0.0 turning this feature off and 1.0 corresponding to old enabled behavior. Now values between 0.0 and 1.0 are also possible.
+
 ### Q-Moving-Average
 Tested some variants of Gudmundsson and Björnsson and Feldman and Domshlak. For a description see: