📦 NEW: Add policy temperature

docker-compose.yml

@@ -39,6 +39,25 @@ services:
       - ./builds/princhess-main:/engines/princhess-main
       - ./syzygy:/syzygy:ro
 
+  sprt_gain_5k:
+    build:
+      context: .
+      dockerfile: bin/Dockerfile.cutechess-cli
+    command:
+      -engine cmd=/engines/princhess name=princhess
+      -engine cmd=/engines/princhess-main name=princhess-main
+
+      -each proto=uci tc=inf nodes=5000
+            option.SyzygyPath=/syzygy option.Hash=128 option.Threads=1
+      -sprt elo0=0 elo1=5 alpha=0.05 beta=0.1
+      -openings file=/books/4moves_noob.epd format=epd order=random
+      -games 2 -repeat -rounds 7500
+      -recover -ratinginterval 10 -concurrency 6
+    volumes:
+      - ./target/release/princhess:/engines/princhess
+      - ./builds/princhess-main:/engines/princhess-main
+      - ./syzygy:/syzygy:ro
+
   sprt_gain_ltc:
     build:
       context: .
@@ -141,12 +160,12 @@ services:
       -engine cmd=/engines/princhess-main name=princhess-main
       -engine cmd=/engines/princhess name=princhess- option.
 
-      -each proto=uci tc=8+0.08
+      -each proto=uci tc=inf nodes=5000
             option.SyzygyPath=/syzygy option.Hash=128 option.Threads=1
       -tournament gauntlet
       -openings file=/books/4moves_noob.epd format=epd order=random
-      -games 2 -repeat -rounds 50
-      -recover -ratinginterval 10 -concurrency 6
+      -games 2 -repeat -rounds 250
+      -recover -ratinginterval 100 -concurrency 6
     volumes:
       - ./target/release/princhess:/engines/princhess
       - ./builds/princhess-main:/engines/princhess-main

src/evaluation.rs

@@ -71,8 +71,8 @@ pub fn evaluate_state_flag(state: &State, moves: &MoveList) -> Flag {
     state.side_to_move().fold_wb(flag, flag.flip())
 }
 
-pub fn evaluate_policy(state: &State, moves: &MoveList) -> Vec<f32> {
-    run_policy_net(state, moves)
+pub fn evaluate_policy(state: &State, moves: &MoveList, t: f32) -> Vec<f32> {
+    run_policy_net(state, moves, t)
 }
 
 const QAB: f32 = 256. * 256.;
@@ -123,7 +123,7 @@ fn run_eval_net(state: &State) -> f32 {
     (result as f32 / QAB).tanh()
 }
 
-fn run_policy_net(state: &State, moves: &MoveList) -> Vec<f32> {
+fn run_policy_net(state: &State, moves: &MoveList, t: f32) -> Vec<f32> {
     let mut evalns = Vec::with_capacity(moves.len());
 
     if moves.is_empty() {
@@ -143,7 +143,7 @@ fn run_policy_net(state: &State, moves: &MoveList) -> Vec<f32> {
         }
     });
 
-    math::softmax(&mut evalns);
+    math::softmax(&mut evalns, t);
 
     evalns
 }

src/math.rs

@@ -1,9 +1,9 @@
-pub fn softmax(arr: &mut [f32]) {
+pub fn softmax(arr: &mut [f32], t: f32) {
     let max = max(arr);
     let mut s = 0.;
 
     for x in &mut *arr {
-        *x = fastapprox::faster::exp(*x - max);
+        *x = fastapprox::faster::exp((*x - max) / t);
         s += *x;
     }
     for x in &mut *arr {

src/mcts.rs

@@ -4,6 +4,7 @@ use std::sync::Arc;
 use std::thread::JoinHandle;
 
 use crate::evaluation;
+use crate::options::get_policy_temperature;
 use crate::search::{TimeManagement, SCALE};
 pub use crate::search_tree::*;
 use crate::state::State;
@@ -123,7 +124,8 @@ impl Mcts {
         let root_moves = root_node.hots();
 
         let state_moves = root_state.available_moves();
-        let state_moves_eval = evaluation::evaluate_policy(root_state, &state_moves);
+        let state_moves_eval =
+            evaluation::evaluate_policy(root_state, &state_moves, get_policy_temperature());
 
         let mut moves: Vec<(&HotMoveInfo, f32)> = root_moves.iter().zip(state_moves_eval).collect();
         moves.sort_by_key(|(h, e)| (h.average_reward().unwrap_or(*e) * SCALE) as i64);

src/options.rs

@@ -8,6 +8,7 @@ static HASH_SIZE_MB: AtomicUsize = AtomicUsize::new(16);
 
 static CPUCT: Lazy<RwLock<f32>> = Lazy::new(|| RwLock::new(1.85));
 static CVISITS_SELECTION: Lazy<RwLock<f32>> = Lazy::new(|| RwLock::new(0.01));
+static POLICY_TEMPERATURE: Lazy<RwLock<f32>> = Lazy::new(|| RwLock::new(1.2));
 
 static CHESS960: AtomicBool = AtomicBool::new(false);
 
@@ -47,6 +48,16 @@ pub fn get_cvisits_selection() -> f32 {
     *cv
 }
 
+pub fn set_policy_temperature(t: f32) {
+    let mut pt = POLICY_TEMPERATURE.write().unwrap();
+    *pt = t;
+}
+
+pub fn get_policy_temperature() -> f32 {
+    let pt = POLICY_TEMPERATURE.read().unwrap();
+    *pt
+}
+
 pub fn set_chess960(c: bool) {
     CHESS960.store(c, Ordering::Relaxed);
 }

src/search_tree.rs

@@ -9,7 +9,7 @@ use crate::arena::Error as ArenaError;
 use crate::evaluation::{self, Flag};
 use crate::math;
 use crate::mcts::{eval_in_cp, ThreadData};
-use crate::options::{get_cpuct, get_cvisits_selection};
+use crate::options::{get_cpuct, get_cvisits_selection, get_policy_temperature};
 use crate::search::{to_uci, TimeManagement, SCALE};
 use crate::state::State;
 use crate::transposition_table::{LRAllocator, LRTable, TranspositionTable};
@@ -26,6 +26,7 @@ pub struct SearchTree {
     root_state: State,
 
     cpuct: f32,
+    policy_t: f32,
 
     #[allow(dead_code)]
     root_table: TranspositionTable,
@@ -162,14 +163,15 @@ fn create_node<'a, F>(
     state: &State,
     tb_hits: &AtomicUsize,
     alloc_slice: F,
+    policy_t: f32,
 ) -> Result<SearchNode, ArenaError>
 where
     F: FnOnce(usize) -> Result<&'a mut [HotMoveInfo], ArenaError>,
 {
     let moves = state.available_moves();
 
     let state_flag = evaluation::evaluate_state_flag(state, &moves);
-    let move_eval = evaluation::evaluate_policy(state, &moves);
+    let move_eval = evaluation::evaluate_policy(state, &moves, policy_t);
 
     if state_flag.is_tablebase() {
         tb_hits.fetch_add(1, Ordering::Relaxed);
@@ -195,9 +197,12 @@ impl SearchTree {
 
         let root_table = TranspositionTable::for_root();
 
-        let mut root_node = create_node(&state, &tb_hits, |sz| {
-            root_table.arena().allocator().alloc_slice(sz)
-        })
+        let mut root_node = create_node(
+            &state,
+            &tb_hits,
+            |sz| root_table.arena().allocator().alloc_slice(sz),
+            1.0,
+        )
         .expect("Unable to create root node");
 
         previous_table.lookup_into(&state, &mut root_node);
@@ -209,7 +214,7 @@ impl SearchTree {
                 .map(|m| m.average_reward().unwrap_or(-SCALE) / SCALE)
                 .collect();
 
-            math::softmax(&mut avg_rewards);
+            math::softmax(&mut avg_rewards, 1.0);
 
             root_node.update_policy(&avg_rewards);
         }
@@ -218,6 +223,7 @@ impl SearchTree {
             root_state: state,
             root_node,
             cpuct: get_cpuct(),
+            policy_t: get_policy_temperature(),
             root_table,
             ttable: LRTable::new(current_table, previous_table),
             num_nodes: 1.into(),
@@ -387,8 +393,12 @@ impl SearchTree {
             };
         }
 
-        let mut created_here =
-            create_node(state, &self.tb_hits, |sz| tld.allocator.alloc_move_info(sz))?;
+        let mut created_here = create_node(
+            state,
+            &self.tb_hits,
+            |sz| tld.allocator.alloc_move_info(sz),
+            self.policy_t,
+        )?;
 
         self.ttable.lookup_into(state, &mut created_here);
 

src/uci.rs

@@ -5,6 +5,7 @@ use std::thread;
 
 use crate::options::{
     set_chess960, set_cpuct, set_cvisits_selection, set_hash_size_mb, set_num_threads,
+    set_policy_temperature,
 };
 use crate::search::Search;
 use crate::search_tree::print_size_list;
@@ -80,6 +81,7 @@ pub fn uci() {
     println!("option name SyzygyPath type string default <empty>");
     println!("option name CPuct type string default 1.85");
     println!("option name CVisitsSelection type string default 0.01");
+    println!("option name PolicyTemperature type string default 1.2");
     println!("option name UCI_Chess960 type check default false");
 
     println!("uciok");
@@ -140,6 +142,7 @@ impl UciOption {
             "hash" => self.set_option(set_hash_size_mb),
             "cpuct" => self.set_option(set_cpuct),
             "cvisitsselection" => self.set_option(set_cvisits_selection),
+            "policytemperature" => self.set_option(set_policy_temperature),
             "uci_chess960" => self.set_option(set_chess960),
             _ => warn!("Badly formatted or unknown option"),
         }