cfr 0.6.0

Counterfactual regret minimization solver for two-player zero-sum incomplete-information games
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
mod gambit;
mod json;

use cfr::{Game, GameError, GameTree, PlayerNum, RegretParams, SolveMethod, SolveParams};
use gambit::{GambitError, GambitInfoset, GambitNode};
use gambit_parser::ExtensiveFormGame;
use clap::{Parser, ValueEnum};
use serde::Serialize;
use std::borrow::Borrow;
use std::collections::HashMap;
use std::fmt::{self, Display, Formatter};
use std::fs::File;
use std::hash::Hash;
use std::io;
use std::io::{BufReader, Read};

#[derive(Clone, Copy, Debug, PartialEq, Eq, ValueEnum)]
enum Method {
    /// No sampling
    Full,
    /// Sample chance nodes
    Sampled,
    /// Sample chance and the other player
    External,
}

#[derive(Clone, Copy, Debug, PartialEq, Eq, ValueEnum)]
enum InputFormat {
    /// Auto-detect format from file extension and contents
    Auto,
    /// Gambit style `.efg` format
    Gambit,
    /// Json game dsl: https://github.com/erikbrinkman/cfr#json-format
    Json,
}

#[derive(Clone, Copy, Debug, PartialEq, Eq, ValueEnum)]
enum Discount {
    /// No discounting
    Vanilla,
    /// Linear discounting of regret and strategies
    Lcfr,
    /// Forget negative regrets and quadratic strategy discounting (CFR+)
    CfrPlus,
    /// Rough average of LCFR and CFR+
    Dcfr,
    /// DCFR modified to prune poor actions
    DcfrPrune,
}

impl Discount {
    fn into_params(self) -> RegretParams {
        match self {
            Discount::Vanilla => RegretParams::vanilla(),
            Discount::Lcfr => RegretParams::lcfr(),
            Discount::CfrPlus => RegretParams::cfr_plus(),
            Discount::Dcfr => RegretParams::dcfr(),
            Discount::DcfrPrune => RegretParams::dcfr_prune(),
        }
    }
}

/// Counterfactual regret minimization solver for two-player zero-sum incomplete-information games
///
/// This program reads a two-player zero-sum perfect-recall extensive-form game in several formats
/// and then uses veriants of counterfactual regret minimization to find approximate nash
/// equilibria. The result will be a json object like:
///
/// `{ "expected_one_utility": <number> , "player_one_regret": <number>, "player_two_regret":
/// <number>, "regret": <number>, "player_one_strategy": <strat>, "player_two_strategy": <strat>
/// }`
///
/// where strats will be mapping from infosets, to actions to probabilities. Zero probability
/// actions will be omitted.
///
/// For more information see: https://github.com/erikbrinkman/cfr
#[derive(Parser, Debug)]
#[clap(author, version, about)]
struct Args {
    /// Try pruning strategies played less than this fraction
    ///
    /// After finding a solution subject to max-regret and max-iters criteria, this will try
    /// pruning any strategy played less than this fraction of the time. If pruning all strategies
    /// below this threshold produces less regret then the pruned strategy is output.
    #[clap(short, long, value_parser, default_value_t = 0.0)]
    clip_threshold: f64,

    /// Terminate solving early if regret is below `max_regret`
    #[clap(short = 'r', long, value_parser, default_value_t = 0.0)]
    max_regret: f64,

    /// Stop after `max_iters`
    #[clap(short = 't', long, value_parser, default_value_t = 1000)]
    max_iters: u64,

    /// Amount of parallelism to use for solving
    ///
    /// If set to zero (default), this will use rusts `std::thread::available_parallelism` falling
    /// back to single threaded if this can't determine an amount.
    #[clap(short, long, value_parser, default_value_t = 0)]
    parallel: usize,

    /// Seed for deterministic outcome sampling
    ///
    /// Sampling is fully deterministic given this seed, so repeated runs with the same seed produce
    /// identical strategies. Vary it to get independent samples.
    #[clap(short, long, value_parser, default_value_t = 0)]
    seed: u64,

    /// Method to use for game solving
    #[clap(short, long, value_enum, default_value_t = Method::External)]
    method: Method,

    /// Format of the input game file
    #[clap(long, value_enum, default_value_t = InputFormat::Auto)]
    input_format: InputFormat,

    /// Discounted CFR parameters
    #[clap(short, long, value_enum, default_value_t = Discount::Dcfr)]
    discount: Discount,

    /// Read game from a file instead of from stdin
    #[clap(short, long, value_parser, default_value = "-")]
    input: String,

    /// Write results to a file instead of stdout
    #[clap(short, long, value_parser, default_value = "-")]
    output: String,

    /// Key gambit infosets by their always-unique number instead of name
    ///
    /// The default keys by name, which errors when distinct infosets share one.
    #[clap(long, value_parser)]
    infoset_numbers: bool,
}

#[derive(Debug, Serialize)]
struct Strategy(HashMap<String, HashMap<String, f64>>);

impl Strategy {
    /// Collect a named strategy into a JSON-serializable map, keying each infoset with `key`.
    ///
    /// Distinct infosets must key to distinct strings -- otherwise one would silently overwrite the
    /// other in the output object -- so a collision is surfaced as [`CliError::DuplicateInfoset`]
    /// rather than dropped or panicked on.
    fn from_named<S, A, T, N>(
        named_strategies: impl IntoIterator<Item = (S, A)>,
        key: impl Fn(S) -> String,
    ) -> Result<Self, CliError>
    where
        A: IntoIterator<Item = (T, N)>,
        T: Display,
        N: Borrow<f64>,
    {
        let mut map = HashMap::new();
        for (info, actions) in named_strategies {
            let name = key(info);
            let probs = actions
                .into_iter()
                .filter(|(_, p)| p.borrow() > &0.0)
                .map(|(a, p)| (a.to_string(), *p.borrow()))
                .collect();
            if map.insert(name.clone(), probs).is_some() {
                return Err(CliError::DuplicateInfoset(name));
            }
        }
        Ok(Strategy(map))
    }
}

#[derive(Debug, Serialize)]
struct Output {
    regret: f64,
    player_one_utility: f64,
    player_two_utility: f64,
    player_one_regret: f64,
    player_two_regret: f64,
    player_one_strategy: Strategy,
    player_two_strategy: Strategy,
}

fn main() {
    let args = Args::parse();
    let mut buff = String::new();
    if args.input == "-" {
        io::stdin().lock().read_to_string(&mut buff).unwrap();
    } else {
        BufReader::new(File::open(&args.input).unwrap())
            .read_to_string(&mut buff)
            .unwrap();
    }
    let out = match args.input_format {
        InputFormat::Json => solve_json(&buff, &args),
        InputFormat::Gambit => solve_gambit(&buff, &args),
        InputFormat::Auto if args.input.ends_with(".json") => solve_json(&buff, &args),
        InputFormat::Auto if args.input.ends_with(".efg") => solve_gambit(&buff, &args),
        InputFormat::Auto => solve_auto(&buff, &args),
    }
    .unwrap();
    if args.output == "-" {
        serde_json::to_writer(io::stdout(), &out).unwrap();
    } else {
        serde_json::to_writer(File::create(args.output).unwrap(), &out).unwrap();
    };
}

/// A user-facing failure from reading or solving the input game.
enum CliError {
    /// The input wasn't valid json; carries serde_json's line/column detail.
    ParseJson(serde_json::Error),
    /// The input wasn't a valid gambit `.efg` file; carries the parser's rendered message (which
    /// borrows the input, so it's captured as an owned string).
    ParseGambit(String),
    /// The gambit game couldn't be read as a two-player zero-sum game.
    Gambit(GambitError),
    /// The game violated a structural invariant the solver requires.
    Materialize(GameError),
    /// Two distinct infosets render to the same output name (carries the shared name).
    DuplicateInfoset(String),
    /// Auto-detection matched no known format.
    UnknownFormat,
}

impl Display for CliError {
    fn fmt(&self, out: &mut Formatter<'_>) -> fmt::Result {
        match self {
            CliError::ParseJson(err) => write!(
                out,
                "couldn't parse json game definition: {err} : https://github.com/erikbrinkman/cfr#json-error"
            ),
            CliError::ParseGambit(err) => write!(
                out,
                "couldn't parse gambit game definition: {err} : https://github.com/erikbrinkman/cfr#gambit-error"
            ),
            CliError::Gambit(err) => Display::fmt(err, out),
            CliError::Materialize(err) => write!(
                out,
                "couldn't extract a compact game representation due to problems with the structure ({err}) : https://github.com/erikbrinkman/cfr#game-error"
            ),
            CliError::DuplicateInfoset(name) => write!(
                out,
                "two distinct infosets share the name {name:?}; pass `--infoset-numbers` to key gambit infosets by number instead : https://github.com/erikbrinkman/cfr#duplicate-infosets"
            ),
            CliError::UnknownFormat => write!(
                out,
                "couldn't parse any known format; try specifying your format with `--input-format` : https://github.com/erikbrinkman/cfr#auto-error"
            ),
        }
    }
}

// surface the Display message, not the variant, when `main` unwraps a failure
impl fmt::Debug for CliError {
    fn fmt(&self, out: &mut Formatter<'_>) -> fmt::Result {
        Display::fmt(self, out)
    }
}

/// Parse a json game (the `&State` is the game) and solve it. Json infosets are uniquely named, so
/// they're always keyed by name.
fn solve_json(buff: &str, args: &Args) -> Result<Output, CliError> {
    let state: json::State = serde_json::from_str(buff).map_err(CliError::ParseJson)?;
    solve_game(&state, 0.0, args, |info| info.to_string())
}

/// Solve a parsed gambit game, keying infosets by number or name per `--infoset-numbers`.
fn solve_gambit_game(game: GambitNode<'_, '_>, sum: f64, args: &Args) -> Result<Output, CliError> {
    if args.infoset_numbers {
        solve_game(game, sum, args, |info: &GambitInfoset<'_>| {
            info.number().to_string()
        })
    } else {
        solve_game(game, sum, args, |info: &GambitInfoset<'_>| info.to_string())
    }
}

/// Parse a gambit game (the [`GambitNode`] cursor keeps borrowing the parsed tree) and solve it.
fn solve_gambit(buff: &str, args: &Args) -> Result<Output, CliError> {
    let parsed =
        ExtensiveFormGame::try_from(buff).map_err(|err| CliError::ParseGambit(err.to_string()))?;
    let game = GambitNode::try_from(&parsed).map_err(CliError::Gambit)?;
    let sum = game.sum();
    solve_gambit_game(game, sum, args)
}

/// Try each known input format in turn.
fn solve_auto(buff: &str, args: &Args) -> Result<Output, CliError> {
    if let Ok(state) = serde_json::from_str::<json::State>(buff) {
        solve_game(&state, 0.0, args, |info| info.to_string())
    } else if let Ok(parsed) = ExtensiveFormGame::try_from(buff) {
        let game = GambitNode::try_from(&parsed).map_err(CliError::Gambit)?;
        let sum = game.sum();
        solve_gambit_game(game, sum, args)
    } else {
        Err(CliError::UnknownFormat)
    }
}

/// Materialize a [`Game`] into a tree, solve it, and format the result. `sum` offsets the reported
/// utilities by the game's constant-sum value.
fn solve_game<G>(
    game: G,
    sum: f64,
    args: &Args,
    infoset_key: impl Fn(&G::Infoset) -> String,
) -> Result<Output, CliError>
where
    G: Game,
    G::Infoset: Eq + Hash,
    G::Action: Eq + Hash + Display,
    G::ChanceInfoset: Eq + Hash,
{
    let game = GameTree::from_game(game).map_err(CliError::Materialize)?;
    let max_iters = if args.max_iters == 0 {
        u64::MAX
    } else {
        args.max_iters
    };
    let method = match args.method {
        Method::Full => SolveMethod::Full,
        Method::Sampled => SolveMethod::Sampled,
        Method::External => SolveMethod::External,
    };
    let (mut strategies, _) = game
        .solve(
            method,
            max_iters,
            args.max_regret,
            args.parallel,
            SolveParams {
                regret: args.discount.into_params(),
                seed: args.seed,
                ..Default::default()
            },
        )
        .unwrap();
    // build the unpruned strategy (converting labels to owned strings) before truncating in place,
    // so we never have to clone `Strategies`; only rebuild from the pruned tree if it's better
    let info = strategies.get_info();
    let [one, two] = strategies.as_named();
    let unpruned = [
        Strategy::from_named(one, &infoset_key)?,
        Strategy::from_named(two, &infoset_key)?,
    ];
    strategies.truncate(args.clip_threshold);
    let pruned_info = strategies.get_info();
    let (info, [player_one_strategy, player_two_strategy]) =
        if pruned_info.regret() < info.regret() {
            let [one, two] = strategies.as_named();
            (
                pruned_info,
                [
                    Strategy::from_named(one, &infoset_key)?,
                    Strategy::from_named(two, &infoset_key)?,
                ],
            )
        } else {
            (info, unpruned)
        };
    Ok(Output {
        regret: info.regret(),
        player_one_utility: info.player_utility(PlayerNum::One) + sum,
        player_two_utility: info.player_utility(PlayerNum::Two) + sum,
        player_one_regret: info.player_regret(PlayerNum::One),
        player_two_regret: info.player_regret(PlayerNum::Two),
        player_one_strategy,
        player_two_strategy,
    })
}

#[cfg(test)]
mod tests {
    use super::{Args, CliError, GameError, GambitError, solve_auto, solve_gambit, solve_json};
    use clap::{CommandFactory, Parser};

    // matching pennies as a json game: a converging two-player game to drive the solve path
    const PENNIES: &str = r#"{ "player": { "player_one": true, "infoset": "p1", "actions": {
        "h": { "player": { "player_one": false, "infoset": "p2", "actions": {
            "h": { "terminal": 1.0 }, "t": { "terminal": -1.0 } } } },
        "t": { "player": { "player_one": false, "infoset": "p2", "actions": {
            "h": { "terminal": -1.0 }, "t": { "terminal": 1.0 } } } } } } }"#;

    fn default_args() -> Args {
        Args::parse_from(["cfr"])
    }

    #[test]
    fn test_cli() {
        Args::command().debug_assert();
    }

    #[test]
    fn auto_detects_json() {
        solve_auto(r#"{ "terminal": 0.0 }"#, &default_args()).unwrap();
    }

    #[test]
    fn auto_detects_gambit() {
        solve_auto(r#"EFG 2 R "" { "" "" } t "" 1 "" { 0 0 }"#, &default_args()).unwrap();
    }

    #[test]
    fn auto_rejects_unknown() {
        let err = solve_auto("random", &default_args()).unwrap_err();
        assert!(matches!(err, CliError::UnknownFormat));
    }

    #[test]
    fn error_messages_render() {
        let json_err = serde_json::from_str::<serde_json::Value>("not json").unwrap_err();
        let errors = [
            CliError::ParseJson(json_err),
            CliError::ParseGambit("error parsing game at: 'oops'".to_owned()),
            CliError::Gambit(GambitError::NotConstantSum),
            CliError::Materialize(GameError::ImperfectRecall),
            CliError::DuplicateInfoset("dup".to_owned()),
            CliError::UnknownFormat,
        ];
        for err in errors {
            assert!(!err.to_string().is_empty());
            // Debug delegates to Display, so the message surfaces when `main` unwraps
            assert_eq!(format!("{err:?}"), err.to_string());
        }
    }

    #[test]
    fn reports_input_errors() {
        let args = default_args();
        assert!(matches!(
            solve_json("not json", &args),
            Err(CliError::ParseJson(_))
        ));
        assert!(matches!(
            solve_gambit("not gambit", &args),
            Err(CliError::ParseGambit(_))
        ));
        // parses as gambit but has three players
        assert!(matches!(
            solve_gambit(r#"EFG 2 R "" { "" "" "" } t "" 1 "" { 0 0 0 }"#, &args),
            Err(CliError::Gambit(_))
        ));
        // parses as json but isn't a valid game (a zero-probability lone chance outcome)
        let bad = r#"{ "chance": { "outcomes": { "a": { "prob": 0.0, "state": { "terminal": 0.0 } } } } }"#;
        assert!(matches!(
            solve_json(bad, &args),
            Err(CliError::Materialize(_))
        ));
    }

    #[test]
    fn solves_across_methods_and_discounts() {
        for method in ["full", "sampled", "external"] {
            for discount in ["vanilla", "lcfr", "cfr-plus", "dcfr", "dcfr-prune"] {
                let args = Args::parse_from(["cfr", "-m", method, "-d", discount, "-t", "50"]);
                let out = solve_json(PENNIES, &args).unwrap();
                assert!(out.regret.is_finite());
            }
        }
    }

    // two distinct player-two infosets that share a name (here both unnamed, so both empty): keyed by
    // name they collide, keyed by number they don't
    const SHARED_NAME_GAME: &str = r#"EFG 2 R "" { "" "" } p "" 1 1 "" { "a" "b" } 0 p "" 2 1 "" { "x" "y" } 0 t "" 1 "" { 1 -1 } t "" 2 "" { -1 1 } p "" 2 2 "" { "x" "y" } 0 t "" 3 "" { -1 1 } t "" 4 "" { 1 -1 }"#;

    #[test]
    fn duplicate_infoset_names_error_by_default() {
        // two distinct infosets sharing a name collide in the output map, so the CLI errors rather than
        // dropping one (this also covers unnamed infosets, whose name is the empty string)
        assert!(matches!(
            solve_gambit(SHARED_NAME_GAME, &default_args()),
            Err(CliError::DuplicateInfoset(_))
        ));
    }

    #[test]
    fn infoset_numbers_flag_resolves_collisions() {
        // keying by the (always unique) infoset number lets the same game solve
        let numbered = Args::parse_from(["cfr", "--infoset-numbers"]);
        let out = solve_gambit(SHARED_NAME_GAME, &numbered).unwrap();
        assert_eq!(out.player_two_strategy.0.len(), 2, "both p2 infosets present");
    }

    #[test]
    fn reports_both_constant_sum_utilities() {
        // a constant-sum-4 game (no decisions): chance pays player one 3 or 1 with equal odds, so
        // player one expects 2.0 and -- since payoffs sum to 4 -- player two also expects 2.0. The
        // constant-sum offset must be added back to *both* reported utilities.
        let game = r#"EFG 2 R "" { "" "" } c "" 1 "c" { "x" 1/2 "y" 1/2 } 0 t "" 1 "" { 3 1 } t "" 2 "" { 1 3 }"#;
        let out = solve_gambit(game, &default_args()).unwrap();
        assert!(
            (out.player_one_utility - 2.0).abs() < 1e-9,
            "player one: {}",
            out.player_one_utility
        );
        assert!(
            (out.player_two_utility - 2.0).abs() < 1e-9,
            "player two: {}",
            out.player_two_utility
        );
    }

    #[test]
    fn clip_threshold_runs_the_pruning_path() {
        let args = Args::parse_from(["cfr", "-c", "0.1", "-t", "200"]);
        let out = solve_json(PENNIES, &args).unwrap();
        assert!(out.regret.is_finite());
    }
}