Merge ead4219 into bfd70fc

CMakeLists.txt

@@ -10,7 +10,7 @@
 cmake_minimum_required(VERSION 3.0)
 
 # Set project name here.
-project(Goldfish VERSION 1.6.0 LANGUAGES CXX)
+project(Goldfish VERSION 1.7.0 LANGUAGES CXX)
 
 
 # Include stuff. No change needed.

README.md

@@ -32,6 +32,9 @@ original author for his great work. Starting with this project meant a stable
 starting point, with all the rules sorted out, basic search in place and a test
 suite to make sure it all works as expected.
 
+At this point, the newest version of Goldfish has been substantially revamped and improved 
+in most aspects (see [Road map](#road-map)).  
+
 ## Why Goldfish?
 
 For some reason, several top chess engines have names of different fish, e.g.
@@ -40,25 +43,36 @@ and so it seemed only fitting for my somewhat limited program to be named this.
 
 ## Road map
 
-The current plan for the project is to improve the strength, including, but not
-limited to:
+The current plan for the project is to improve the strength. The following is a
+non-exhaustive list of possibilities for future additions, including all features that have
+been added so far. The list is inspired in large part by [this writeup](http://www.frayn.net/beowulf/theory.html).
 
 - [X] Making the engine playable on [lichess.org](lichess.org)
 - [X] Complete refactoring of base types
 - [X] Null move pruning
 - [X] Transposition table
 - [X] Check extensions
 - [X] Killer move heuristic
-- [ ] Passed pawn
+- [X] Principal variation search
+- [ ] Internal iterative deepening
+- [ ] Aspiration window search
+- [ ] Futility pruning
 - [ ] Staged move generation
-- [ ] Better search algorithms, such as MTD-bi
+- [ ] Better search algorithms, such as MTD-bi (?)
 - [ ] More sophisticated static evaluation
+  + [ ] Extra considerations for passed pawns
+  + [ ] Piece square tables
+  + [ ] King safety
+  + [ ] Center control
+  + [ ] Rooks on the 7th rank
+  + [ ] Bishops on main diagonals
+
 
 Each significant change will result in a new version of the engine (see
-releases). In the following you see a relative rating between the current
-versions. In this rating system, v1.0 is held at 2000 rating points, and the
-others are adjusted accordingly. This gives an impression of the relative
-improvements of the engine over time.
+[releases](https://github.com/bsamseth/Goldfish/releases)). In the following
+you see a relative rating between the current versions. In this rating system,
+v1.0 is held at 2000 rating points, and the others are adjusted accordingly.
+This gives an impression of the relative improvements of the engine over time.
 
    |# |PLAYER     |       RATING | POINTS | PLAYED |  (%)|
    |:---:|---|:---:|:---:|:---:|:---:|

src/search.cpp

@@ -364,7 +364,34 @@ void Search::search_root(Depth depth, int alpha, int beta) {
         protocol.send_status(false, current_depth, current_max_depth, total_nodes, current_move, current_move_number);
 
         position.make_move(move);
-        int value = -search(depth - 1, -beta, -alpha, ply + 1);
+        int value;
+        //
+        // Principal Variation Search
+        //
+        // Search first move fully, then just check for moves that will
+        // improve alpha using a 1-point search window. If first move was
+        // best, then we will save lots of time as bounding is much faster than
+        // finding exact scores. Given a good ordering (which we have due to
+        // iterative deepening) the first move will be very good, and lots of
+        // cutoffs can be made.
+        //
+        // If we find a later move that actually improves alpha, we must search this
+        // properly to find its value. The idea is that this drawback is smaller than
+        // the improvements gained.
+        //
+        if (depth > 2 and i > 0) {
+
+            value = -search(depth - 1, -alpha - 1, -alpha, ply + 1);
+
+            if (value >= alpha) {
+                // PV search failed high, need to do a full search.
+                value = -search(depth - 1, -beta, -alpha, ply + 1);
+            }
+        }
+        // First move, or to shallow for PV search - search fully.
+        else {
+            value = -search(depth - 1, -beta, -alpha, ply + 1);
+        }
         position.undo_move(move);
 
         if (abort) {
@@ -513,7 +540,21 @@ int Search::search(Depth depth, int alpha, int beta, int ply) {
         position.make_move(move);
         if (!position.is_check(~position.active_color)) {
             searched_moves++;
-            value = -search(depth - 1, -beta, -alpha, ply + 1);
+            //
+            // Principal Variation Search (see search_root for details).
+            //
+            if (depth > 1 and i > 0) {
+
+                value = -search(depth - 1, -alpha - 1, -alpha, ply + 1);
+
+                if (value >= alpha) {
+                    // PV search failed high, need to do a full search.
+                    value = -search(depth - 1, -beta, -alpha, ply + 1);
+                }
+            } else {
+                // First move or to shallow - do full search.
+                value = -search(depth - 1, -beta, -alpha, ply + 1);
+            }
         }
         position.undo_move(move);