initial implementation of MoveValueStats

src/elo.rs

@@ -71,7 +71,7 @@ pub fn run() {
             .to_vec()
         })
         .flat_map(move |prev_c| {
-            [ChildrenEvalMethod::MinMax, ChildrenEvalMethod::MinMaxFlat].map(move |method| {
+            [ChildrenEvalMethod::MinMax, ChildrenEvalMethod::MinMaxProb].map(move |method| {
                 FutureMoveConfig {
                     children_eval_method: method,
                     ..prev_c

src/logic/board_value.rs

@@ -33,10 +33,4 @@ impl BoardValueMap {
         ];
         Self(PosMap::from(POSITION_VALUES))
     }
-
-    pub const fn flat() -> Self {
-        Self(PosMap::from(
-            [[1; Board::SIZE as usize]; Board::SIZE as usize],
-        ))
-    }
 }

src/logic/future_moves.rs

@@ -1,3 +1,4 @@
+use super::r#move::{MoveCoord, MoveValueConfig, MoveValueStats};
 use crate::{
     logic::r#move::Move,
     repr::{Board, Piece, Winner},
@@ -15,8 +16,6 @@ use std::{
     },
 };
 
-use super::r#move::MoveCoord;
-
 #[derive(Allocative)]
 pub struct FutureMoves {
     /// Arena containing all [`Move`]
@@ -25,9 +24,6 @@ pub struct FutureMoves {
     /// Index of the [`Move`] tree's root node
     current_root: Option<usize>,
 
-    /// Current generated depth of the Arena
-    current_depth: usize,
-
     /// Color w.r.t
     agent_color: Piece,
 
@@ -94,15 +90,10 @@ impl std::fmt::Display for FutureMoveConfig {
 #[derive(Debug, Clone, Copy, Allocative)]
 #[allow(dead_code)]
 pub enum ChildrenEvalMethod {
-    Average,
-    AverageDivDepth,
-
-    MinAvgDivDepth,
-
     /// Best so far?
     MinMax,
 
-    MinMaxFlat,
+    MinMaxProb,
 }
 
 impl Default for ChildrenEvalMethod {
@@ -116,7 +107,6 @@ impl FutureMoves {
         Self {
             arena: Vec::new(),
             current_root: None,
-            current_depth: 0,
             agent_color,
             config,
             board: Board::new(),
@@ -145,39 +135,34 @@ impl FutureMoves {
         indexes
     }
 
-    /// Find the current depth of the arena by
-    /// looking at leaf moves and finding the smallest value
-    fn determine_current_depth(&self) -> Option<usize> {
+    /// Return the current depth of the tree
+    fn current_depth(&self) -> usize {
         // leaf_moves is sorted from min to max depth
-        self.leaf_moves().first().map(|&i| self.depth_of(i))
+        self.leaf_moves()
+            .first()
+            .map(|&i| self.depth_of(i))
+            .unwrap_or(0) // handle empty trees
     }
 
     /// Generate children for all children of `nodes`
     /// only `pub` for the sake of benchmarking
     pub fn extend_layers(&mut self) {
-        // recover from partial tree extention
-        if let Some(current_depth) = self.determine_current_depth() {
-            self.current_depth = current_depth;
-        }
+        let mut leafs = self.leaf_moves().into_iter().collect::<Vec<usize>>();
 
-        for _ in self.current_depth..self.config.max_depth {
+        for _ in self.current_depth()..self.config.max_depth {
             let pstyle_inner = if cfg!(test) || !self.config.print {
                 ""
             } else {
                 &format!(
                     "Generating children (depth: {}/{}): ({{pos}}/{{len}}) {{per_sec}}",
-                    self.current_depth + 1,
+                    self.current_depth() + 1,
                     self.config.max_depth
                 )
             };
 
             let allowed_size = self.config.max_arena_size - self.arena.len();
             let curr_size = Arc::new(AtomicUsize::new(0));
-            let got = self
-                .leaf_moves()
-                .into_iter()
-                .filter(|&i| self.depth_of(i) == self.current_depth)
-                .collect::<Vec<usize>>()
+            leafs = leafs
                 .into_par_iter()
                 .progress_with_style(ProgressStyle::with_template(pstyle_inner).unwrap())
                 .map(|parent_idx| (parent_idx, self.generate_children_raw(parent_idx)))
@@ -189,21 +174,22 @@ impl FutureMoves {
                         true
                     }
                 })
-                .collect::<Vec<(usize, Vec<Move>)>>();
+                .collect::<Vec<(usize, Vec<Move>)>>()
+                .into_iter()
+                .flat_map(|(parent_idx, moves)| {
+                    let start_idx = self.arena.len();
+                    self.arena.extend(moves);
+
+                    let new_indices = start_idx..self.arena.len();
+                    self.arena[parent_idx].children.extend(new_indices.clone());
+                    new_indices
+                })
+                .collect();
 
             // get total # of generated boards
             let got_len = curr_size.load(Ordering::Acquire);
 
-            got.into_iter().for_each(|(parent_idx, moves)| {
-                let start_idx = self.arena.len();
-                self.arena.extend(moves);
-
-                let new_indices = start_idx..self.arena.len();
-                self.arena[parent_idx].children.extend(new_indices);
-            });
-
             self.prune_bad_children();
-            self.current_depth += 1;
             if got_len == allowed_size {
                 // arena has hit the upper limit of size permitted
                 break;
@@ -228,16 +214,7 @@ impl FutureMoves {
     }
 
     fn create_move(&self, coord: MoveCoord, board: Board, color: Piece) -> Move {
-        Move::new(
-            coord,
-            board,
-            color,
-            self.agent_color,
-            !matches!(
-                self.config.children_eval_method,
-                ChildrenEvalMethod::MinMaxFlat
-            ),
-        )
+        Move::new(coord, board, color, self.agent_color, MoveValueConfig {})
     }
 
     fn generate_children_raw(&self, parent_idx: usize) -> Vec<Move> {
@@ -313,7 +290,7 @@ impl FutureMoves {
         let by_depth_vec = self.by_depth(indexes);
 
         // reversed so we build up the value of the closest (in time) moves from the future
-        for (depth, nodes) in by_depth_vec.into_iter().rev() {
+        for (_depth, nodes) in by_depth_vec.into_iter().rev() {
             for idx in nodes {
                 let children_values = self.arena[idx]
                     .children
@@ -321,56 +298,63 @@ impl FutureMoves {
                     .map(|&child| self.arena[child].value)
                     .collect::<Vec<_>>();
 
-                let children_value = match self.config.children_eval_method {
-                    ChildrenEvalMethod::Average => children_values
-                        .into_iter()
-                        .sum::<i32>()
-                        .checked_div(self.arena[idx].children.len() as i32),
-
-                    ChildrenEvalMethod::AverageDivDepth => children_values
-                        .into_iter()
-                        .sum::<i32>()
-                        .checked_div(self.arena[idx].children.len() as i32)
-                        .and_then(|x| x.checked_div(depth as i32)),
-                    ChildrenEvalMethod::MinAvgDivDepth => {
-                        if self.arena[idx].color == self.agent_color {
+                match self.config.children_eval_method {
+                    ChildrenEvalMethod::MinMax => {
+                        let children_value = if self.arena[idx].color == self.agent_color {
                             // get best (for the adversary) enemy play
                             // this assumes the adversary is playing optimally
 
-                            children_values.into_iter().min()
+                            children_values
+                                .into_iter()
+                                .min_by_key(|x| x.value)
+                                .map(|x| x.value)
                         } else {
                             children_values
                                 .into_iter()
-                                .sum::<i32>()
-                                .checked_div(self.arena[idx].children.len() as i32)
-                                .and_then(|x| x.checked_div(depth as i32))
+                                .max_by_key(|x| x.value)
+                                .map(|x| x.value)
                         }
+                        .unwrap_or(0);
+                        // we use `depth` and divided `self_value` by it, idk if this is worth it
+                        // we should really setup some sort of ELO rating for each commit, playing them against
+                        // each other or something, could be cool to benchmark these more subjective things, not
+                        // just performance (cycles/time wise)
+                        self.arena[idx].value.value =
+                            self.arena[idx].self_value as i32 + children_value;
                     }
-
-                    ChildrenEvalMethod::MinMax | ChildrenEvalMethod::MinMaxFlat => {
-                        if self.arena[idx].color == self.agent_color {
+                    ChildrenEvalMethod::MinMaxProb => {
+                        let children_value = if self.arena[idx].color == self.agent_color {
                             // get best (for the adversary) enemy play
                             // this assumes the adversary is playing optimally
 
-                            children_values.into_iter().min()
+                            children_values.iter().min()
                         } else {
-                            children_values.into_iter().max()
+                            children_values.iter().max()
                         }
+                        .cloned()
+                        .unwrap_or(Default::default());
+                        // we use `depth` and divided `self_value` by it, idk if this is worth it
+                        // we should really setup some sort of ELO rating for each commit, playing them against
+                        // each other or something, could be cool to benchmark these more subjective things, not
+                        // just performance (cycles/time wise)
+
+                        let wins = children_values.iter().map(|x| x.wins).sum();
+                        let losses = children_values.iter().map(|x| x.losses).sum();
+
+                        let final_value = MoveValueStats {
+                            wins,
+                            losses,
+                            value: self.arena[idx].self_value as i32 + children_value.value,
+                        };
+                        self.arena[idx].value = final_value;
                     }
                 }
-                .unwrap_or(0);
-
-                // we use `depth` and divided `self_value` by it, idk if this is worth it
-                // we should really setup some sort of ELO rating for each commit, playing them against
-                // each other or something, could be cool to benchmark these more subjective things, not
-                // just performance (cycles/time wise)
-                self.arena[idx].value = self.arena[idx].self_value as i32 + children_value;
             }
         }
     }
 
     fn move_history(&self, idx: usize) -> Option<Vec<(MoveCoord, Piece)>> {
-        if let Some(root) = self.current_root {
+        self.current_root.and_then(|root| {
             let mut hist = Vec::new();
 
             let mut current = Some(idx);
@@ -390,23 +374,19 @@ impl FutureMoves {
             }
 
             Some(hist)
-        } else {
-            None
-        }
+        })
     }
 
     fn get_board_from_idx(&self, idx: usize) -> Option<Board> {
-        if let Some(hist) = self.move_history(idx) {
+        self.move_history(idx).and_then(|hist| {
             let mut board = self.board;
             for (m, c) in hist {
                 if let Some(m) = m {
-                    board.place(m, c).expect("move would not propegate");
+                    board.place(m, c).ok()?;
                 }
             }
             Some(board)
-        } else {
-            None
-        }
+        })
     }
 
     /// Return the best move which is a child of `self.current_root`
@@ -461,7 +441,6 @@ impl FutureMoves {
     fn rebuild_from_board(&mut self, board: Board) {
         self.arena = vec![self.create_move(None, board, !self.agent_color)];
         self.current_root = Some(0);
-        self.current_depth = 0;
         self.board = board;
     }
 
@@ -469,7 +448,6 @@ impl FutureMoves {
         let board = self
             .get_board_from_idx(idx)
             .expect("unable to get board at idx");
-        self.current_depth -= self.depth_of(idx);
         self.current_root = Some(idx);
         self.board = board;
         self.refocus_tree();
@@ -521,7 +499,7 @@ impl FutureMoves {
     }
 
     fn prune_bad_children(&mut self) {
-        if self.current_depth < self.config.min_arena_depth || !self.config.do_prune {
+        if self.current_depth() < self.config.min_arena_depth || !self.config.do_prune {
             return;
         }
 
@@ -530,7 +508,7 @@ impl FutureMoves {
 
         for (depth, indexes) in self.by_depth(0..self.arena.len()) {
             // TODO! maybe update by_depth every iteration or something?
-            if depth > self.current_depth.saturating_sub(self.config.up_to_minus) {
+            if depth > self.current_depth().saturating_sub(self.config.up_to_minus) {
                 return;
             }
 
@@ -580,6 +558,8 @@ impl FutureMoves {
     /// Rebuilds the Arena based on `self.current_root`, prunes unrelated nodes
     fn refocus_tree(&mut self) {
         let Some(root) = self.current_root else {
+            // handle current_root being empty (clear arena and return)
+            self.arena.clear();
             return;
         };
 

src/logic/move.rs

@@ -1,9 +1,48 @@
+use std::cmp::Ordering;
+
 use super::board_value::BoardValueMap;
 use crate::repr::{Board, CoordPair, Piece, Winner};
 use allocative::Allocative;
 
 pub type MoveCoord = Option<CoordPair>;
 
+#[derive(Clone, Copy, Debug, Allocative, PartialEq, Eq, Default)]
+pub struct MoveValueStats {
+    pub wins: u16,
+    pub losses: u16,
+    pub value: i32,
+}
+
+impl MoveValueStats {
+    pub fn chance_win(&self) -> f32 {
+        self.wins as f32 / (self.losses + self.wins) as f32
+    }
+}
+
+impl PartialOrd for MoveValueStats {
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl Ord for MoveValueStats {
+    fn cmp(&self, other: &Self) -> Ordering {
+        if self.wins != 0 || self.losses != 0 || other.wins != 0 || other.losses != 0 {
+            let s_cw = self.chance_win();
+            let o_cw = other.chance_win();
+            if s_cw > o_cw {
+                Ordering::Greater
+            } else if o_cw > s_cw {
+                Ordering::Less
+            } else {
+                Ordering::Equal
+            }
+        } else {
+            self.value.cmp(&other.value)
+        }
+    }
+}
+
 #[derive(Clone, Debug, Allocative)]
 pub struct Move {
     /// Coordinates (i, j) of the move (if it exists)
@@ -23,7 +62,7 @@ pub struct Move {
     pub children: Vec<usize>,
 
     /// Value of this move (including children)
-    pub value: i32,
+    pub value: MoveValueStats,
 
     /// What is the inherit value of this move (not including children)
     pub self_value: i16,
@@ -35,29 +74,35 @@ pub struct Move {
     pub is_trimmed: bool,
 }
 
+pub struct MoveValueConfig {}
+
 impl Move {
     pub fn new(
         coord: MoveCoord,
         board: Board,
         color: Piece,
         agent_color: Piece,
-        use_weighted_bvm: bool,
+        mvc: MoveValueConfig,
     ) -> Self {
         let mut m = Move {
             coord,
             winner: board.game_winner(),
             parent: None,
             children: Vec::new(),
-            value: i32::MIN,
+            value: MoveValueStats {
+                wins: 0,
+                losses: 0,
+                value: 0,
+            },
             color,
             is_trimmed: false,
             self_value: 0,
         };
-        m.self_value = m.compute_self_value(agent_color, &board, use_weighted_bvm);
+        m.self_value = m.compute_self_value(agent_color, &board, mvc);
         m
     }
 
-    fn compute_self_value(&self, agent_color: Piece, board: &Board, use_weighted_bvm: bool) -> i16 {
+    fn compute_self_value(&self, agent_color: Piece, board: &Board, _mvc: MoveValueConfig) -> i16 {
         if self.winner == Winner::Player(!agent_color) {
             // if this board results in the opponent winning, MAJORLY negatively weigh this move
             // NOTE! this branch isn't completely deleted because if so, the bot wouldn't make a move.
@@ -67,19 +112,10 @@ impl Move {
             // results in a win for the agent
             return i16::MAX - 1;
         }
-        // else if self.winner == Winner::Tie {
-        //     // idk what a Tie should be valued?
-        //     return 0;
-        // }
 
         // I guess ignore Ties here, don't give them an explicit value,
-        // because even in the case of ties, we want to have a higher score
 
-        match use_weighted_bvm {
-            true => const { BoardValueMap::weighted() },
-            false => const { BoardValueMap::flat() },
-        }
-        .board_value(board, agent_color)
+        const { BoardValueMap::weighted() }.board_value(board, agent_color)
     }
 
     /// Sort children of the [`Move`] by their self_value in `arena`