titaniumtown/othello
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

MinMax is much better

Browse Source
This commit is contained in:
Simon Gardling 2025-04-03 19:17:37 -04:00
parent a50ca2c1b1
commit cb63b49f7a
Signed by: titaniumtown
GPG Key ID: 9AB28AC10ECE533D
5 changed files with 99 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
benches/future_children.rs
View File

@ -1,6 +1,6 @@
use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
use othello::{
logic::{ChildrenEvalMethod, FutureMoveConfig, FutureMoves},
logic::{FutureMoveConfig, FutureMoves},
repr::{Board, Piece},
};
@ -13,7 +13,7 @@ fn extend_layers_no_pruning(depth: usize) -> usize {
max_arena_size: usize::MAX,
do_prune: false,
print: false,
children_eval_method: ChildrenEvalMethod::AverageDivDepth,
children_eval_method: Default::default(),
};
let mut fut = FutureMoves::new(Piece::Black, config);
fut.update_from_board(&Board::STARTING_POSITION);

4
src/allocs.rs
View File

@ -1,5 +1,5 @@
use crate::{
logic::{ChildrenEvalMethod, FutureMoveConfig, FutureMoves},
logic::{FutureMoveConfig, FutureMoves},
repr::{Board, Piece, Winner},
};
use allocative::FlameGraphBuilder;
@ -18,7 +18,7 @@ pub fn run() {
max_arena_size: 100_000_000,
do_prune: true,
print: true,
children_eval_method: ChildrenEvalMethod::AverageDivDepth,
children_eval_method: Default::default(),
},
);

59
src/elo.rs
View File

@ -1,5 +1,5 @@
use crate::{
agent::Agent,
agent::{Agent, RandomAgent},
complexagent::ComplexAgent,
game_inner::GameInner,
logic::{ChildrenEvalMethod, FutureMoveConfig},
@ -18,7 +18,7 @@ type AgentMaker = Box<dyn Fn(Piece) -> Box<dyn Agent>>;
#[allow(dead_code)]
pub fn run() {
const FMV_BASE: FutureMoveConfig = FutureMoveConfig {
let fmv_base = FutureMoveConfig {
max_depth: 20,
min_arena_depth: 14,
top_k_children: 2,
@ -26,26 +26,26 @@ pub fn run() {
max_arena_size: usize::MAX,
do_prune: false,
print: false,
children_eval_method: ChildrenEvalMethod::AverageDivDepth,
children_eval_method: Default::default(),
};
let configs = [6]
.into_iter()
.map(move |d| FutureMoveConfig {
max_depth: d,
..FMV_BASE
..fmv_base
})
.flat_map(move |prev_c| {
// create children which enable, and disable pruning
[true, false].map(move |do_prune| FutureMoveConfig { do_prune, ..prev_c })
})
.filter(move |move_c| {
if move_c.do_prune {
move_c.max_depth >= 8
} else {
move_c.max_depth < 8
}
[false].map(move |do_prune| FutureMoveConfig { do_prune, ..prev_c })
})
// .filter(move |move_c| {
// if move_c.do_prune {
// move_c.max_depth >= 8
// } else {
// move_c.max_depth < 8
// }
// })
// .flat_map(move |prev_c| {
// [
// ChildrenEvalMethod::Average,
@ -64,12 +64,23 @@ pub fn run() {
}
// different values of top_k_children
[1, 2, 3]
.map(move |top_k_children| FutureMoveConfig {
top_k_children,
..prev_c
})
.to_vec()
[2].map(move |top_k_children| FutureMoveConfig {
top_k_children,
..prev_c
})
.to_vec()
})
.flat_map(move |prev_c| {
[
ChildrenEvalMethod::Average,
ChildrenEvalMethod::AverageDivDepth,
ChildrenEvalMethod::MinAvgDivDepth,
ChildrenEvalMethod::MinMax,
]
.map(move |method| FutureMoveConfig {
children_eval_method: method,
..prev_c
})
})
.flat_map(move |prev_c| {
if !prev_c.do_prune {
@ -79,8 +90,7 @@ pub fn run() {
// different values to be subtracted from max_depth
// to become min_arena_depth
[1, 2, 3]
.into_iter()
[2].into_iter()
.filter(|&x| x <= prev_c.max_depth)
.map(move |ad_offset| FutureMoveConfig {
min_arena_depth: prev_c.max_depth - ad_offset,
@ -95,8 +105,7 @@ pub fn run() {
}
// different values of up_to_minus
[prev_c.max_depth, 1, 2, 3]
.into_iter()
[3].into_iter()
.filter(|&x| x <= prev_c.max_depth)
.map(move |up_to_minus| FutureMoveConfig {
up_to_minus,
@ -105,7 +114,7 @@ pub fn run() {
.collect()
});
let vec: Vec<(String, AgentMaker)> = configs
let mut vec: Vec<(String, AgentMaker)> = configs
.into_iter()
.map(move |config| -> (String, AgentMaker) {
(
@ -114,6 +123,10 @@ pub fn run() {
)
})
.collect();
vec.push((
"RandomAgent".to_string(),
Box::new(move |piece| Box::new(RandomAgent::new(piece))),
));
let mut arena = PlayerArena::new(vec);

77
src/logic/future_moves.rs
View File

@ -79,10 +79,18 @@ impl std::fmt::Display for FutureMoveConfig {
#[allow(dead_code)]
pub enum ChildrenEvalMethod {
Average,
/// AverageDivDepth gives the agent a sense of
/// time when it comes to how far away a potential win or gain
/// is. This performs much better in the Elo Arena than `Average`
AverageDivDepth,
MinAvgDivDepth,
/// Best so far?
MinMax,
}
impl Default for ChildrenEvalMethod {
fn default() -> Self {
Self::MinMax
}
}
impl FutureMoves {
@ -287,6 +295,31 @@ impl FutureMoves {
.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 {
// get best (for the adversary) enemy play
// this assumes the adversary is playing optimally
children_values.into_iter().min()
} 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))
}
}
ChildrenEvalMethod::MinMax => {
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()
} else {
children_values.into_iter().max()
}
}
}
.unwrap_or(0);
@ -571,22 +604,26 @@ impl FutureMoves {
#[cfg(test)]
mod tests {
use std::sync::LazyLock;
use super::*;
const FUTURE_MOVES_CONFIG: FutureMoveConfig = FutureMoveConfig {
max_depth: 3, // we want great-grand children for traversing moves
min_arena_depth: 0,
top_k_children: 1,
up_to_minus: 0,
max_arena_size: 100,
do_prune: false,
print: false,
children_eval_method: ChildrenEvalMethod::AverageDivDepth,
};
static FUTURE_MOVES_CONFIG: LazyLock<FutureMoveConfig> = LazyLock::new(|| {
FutureMoveConfig {
max_depth: 3, // we want great-grand children for traversing moves
min_arena_depth: 0,
top_k_children: 1,
up_to_minus: 0,
max_arena_size: 100,
do_prune: false,
print: false,
children_eval_method: Default::default(),
}
});
#[test]
fn prune_tree_test() {
let mut futm = FutureMoves::new(Piece::Black, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::Black, *FUTURE_MOVES_CONFIG);
futm.update_from_board(&Board::new());
@ -628,7 +665,7 @@ mod tests {
#[test]
fn expand_layer_test() {
let mut futm = FutureMoves::new(Piece::Black, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::Black, *FUTURE_MOVES_CONFIG);
futm.config.max_depth = 1;
futm.update_from_board(&Board::STARTING_POSITION);
@ -653,7 +690,7 @@ mod tests {
#[test]
fn depth_of_test() {
let mut futm = FutureMoves::new(Piece::Black, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::Black, *FUTURE_MOVES_CONFIG);
futm.update_from_board(&Board::new());
@ -681,7 +718,7 @@ mod tests {
#[test]
fn by_depth_test() {
let mut futm = FutureMoves::new(Piece::Black, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::Black, *FUTURE_MOVES_CONFIG);
futm.update_from_board(&Board::new());
@ -707,7 +744,7 @@ mod tests {
/// tests whether or not FutureMoves can recover from multiple skips and then manually regenerating the arena
#[test]
fn skip_move_recovery() {
let mut futm = FutureMoves::new(Piece::Black, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::Black, *FUTURE_MOVES_CONFIG);
let mut board = Board::STARTING_POSITION;
// replay of a test I did
@ -770,7 +807,7 @@ mod tests {
#[test]
fn derive_board() {
let mut futm = FutureMoves::new(Piece::White, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::White, *FUTURE_MOVES_CONFIG);
let mut b = Board::STARTING_POSITION;
futm.update_from_board(&b);
@ -838,7 +875,7 @@ mod tests {
}
}
let mut futm = FutureMoves::new(Piece::White, FUTURE_MOVES_CONFIG);
let mut futm = FutureMoves::new(Piece::White, *FUTURE_MOVES_CONFIG);
futm.update_from_board(&board);
futm.generate();

2
src/main.rs
View File

@ -14,6 +14,8 @@ pub mod repr;
// TODO! make this agent configuration a config option via `clap-rs`
// or maybe even like a TUI menu?
fn main() {
// elo::run();
// return;
let player1 = complexagent::ComplexAgent::new(
Piece::Black,
FutureMoveConfig {