modurl 0.1.0

Modular reinforcement learning algorithms and utilities built on Candle
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
//! Twin Delayed Deep Deterministic Policy Gradient.
//!
//! [`TD3Agent`] extends deterministic replay training with a critic ensemble,
//! target-policy smoothing, and delayed actor updates. [`TD3Logger`] receives
//! replay-update and collection metrics.

use std::fmt::Debug;

use bon::bon;
use candle_core::Tensor;
use candle_nn::{Optimizer, VarMap};

use super::{
    CriticCountRequirement, DeterministicActorCriticAgent,
    DeterministicActorCriticCollectionLogEntry, DeterministicActorCriticError,
    DeterministicActorCriticLogEntry, DeterministicActorCriticLogger,
    DeterministicActorCriticResult, DeterministicActorCriticStrategy, DeterministicCritic,
};
use crate::{
    agents::{
        Agent, ReplayStorageConfig,
        sac::{SACCriticAggregationMode, SACCriticError, aggregate_critic_values},
    },
    gym::MultiGym,
    spaces::{BoxSpace, Space},
};

/// Receives TD3 replay-update and collection metrics.
pub trait TD3Logger<I = ()> {
    /// Records metrics from one replay optimization.
    fn log(&mut self, entry: &DeterministicActorCriticLogEntry);

    /// Records metrics from one vectorized environment step.
    fn log_collection(&mut self, _entry: &DeterministicActorCriticCollectionLogEntry<I>) {}
}

struct TD3LoggingInfo<'a, I> {
    logger: &'a mut dyn TD3Logger<I>,
}

impl<I> DeterministicActorCriticLogger<I> for Option<TD3LoggingInfo<'_, I>> {
    fn log_update(&mut self, entry: &DeterministicActorCriticLogEntry) {
        if let Some(info) = self {
            info.logger.log(entry);
        }
    }

    fn log_collection(&mut self, entry: &DeterministicActorCriticCollectionLogEntry<I>) {
        if let Some(info) = self {
            info.logger.log_collection(entry);
        }
    }
}

pub(crate) struct TD3Strategy {
    target_policy_noise: f64,
    target_noise_clip: f64,
    actor_update_interval: usize,
    target_aggregation_mode: SACCriticAggregationMode,
    actor_aggregation_mode: Option<SACCriticAggregationMode>,
}

impl DeterministicActorCriticStrategy for TD3Strategy {
    fn critic_count_requirement(&self) -> CriticCountRequirement {
        CriticCountRequirement::NonEmpty
    }

    fn actor_update_interval(&self) -> usize {
        self.actor_update_interval
    }

    fn target_policy_noise(&self) -> f64 {
        self.target_policy_noise
    }

    fn target_noise_clip(&self) -> f64 {
        self.target_noise_clip
    }

    /// Takes the elementwise minimum of two target tensors shaped `[batch]`.
    fn aggregate_target_values(&self, values: Vec<Tensor>) -> Result<Tensor, SACCriticError> {
        aggregate_critic_values(&values, self.target_aggregation_mode)
    }

    /// Combines actor-objective Q tensors shaped `[batch_size]` using the
    /// configured mode, or selects the first tensor for canonical TD3.
    fn aggregate_actor_values(&self, values: Vec<Tensor>) -> Result<Tensor, SACCriticError> {
        match self.actor_aggregation_mode {
            Some(mode) => aggregate_critic_values(&values, mode),
            None => values
                .into_iter()
                .next()
                .ok_or(SACCriticError::NoCriticValues),
        }
    }
}

/// Twin Delayed Deep Deterministic Policy Gradient agent.
///
/// The canonical configuration uses two critics and
/// [`SACCriticAggregationMode::Min`] for targets, while optimizing the actor
/// against the first critic. Any non-empty critic ensemble is supported, and
/// target and actor aggregation can be configured independently for
/// experimentation.
pub struct TD3Agent<'a, AO, CO, GE, SE, I = ()>
where
    AO: Optimizer,
    CO: Optimizer,
    GE: Debug,
    SE: Debug,
{
    inner: DeterministicActorCriticAgent<'a, AO, CO, GE, SE, TD3Strategy>,
    logging_info: Option<TD3LoggingInfo<'a, I>>,
}

#[bon]
impl<'a, AO, CO, GE, SE, I> TD3Agent<'a, AO, CO, GE, SE, I>
where
    AO: Optimizer,
    CO: Optimizer,
    GE: Debug,
    SE: Debug,
{
    #[builder]
    /// Builds a TD3 agent and initializes every target network from its online
    /// counterpart.
    pub fn new(
        /// Actor optimized on delayed replay updates.
        #[builder(with = |actor: impl candle_core::Module + 'static| Box::new(actor))]
        online_actor: Box<dyn candle_core::Module>,
        /// Actor used to calculate smoothed next-state Bellman targets.
        #[builder(with = |actor: impl candle_core::Module + 'static| Box::new(actor))]
        target_actor: Box<dyn candle_core::Module>,
        /// Parameters belonging to `online_actor`.
        online_actor_vars: &'a VarMap,
        /// Parameters belonging to `target_actor`.
        target_actor_vars: &'a mut VarMap,
        /// Optimizer for the online actor parameters.
        actor_optimizer: AO,
        /// Non-empty independently optimized critic ensemble.
        critics: Vec<DeterministicCritic<'a, CO>>,
        /// Bounded action space matching the actor output and environment.
        action_space: BoxSpace,
        /// Observation-space contract expected by the actor and critics.
        observation_space: Box<dyn Space<Error = SE>>,
        /// Devices used for replay storage and optimization.
        replay_storage_config: ReplayStorageConfig,
        /// Optional replay-update and collection metric sink.
        logger: Option<&'a mut dyn TD3Logger<I>>,
        /// Bellman discount factor.
        #[builder(default = 0.99)]
        gamma: f64,
        /// Polyak coefficient for actor and critic target updates.
        #[builder(default = 0.005)]
        tau: f64,
        /// Standard deviation of Gaussian collection-action noise.
        #[builder(default = 0.1)]
        exploration_noise: f64,
        /// Standard deviation of Gaussian target-policy noise.
        #[builder(default = 0.2)]
        target_policy_noise: f64,
        /// Componentwise absolute limit for target-policy noise.
        #[builder(default = 0.5)]
        target_noise_clip: f64,
        /// Replay-optimization interval between actor and target updates.
        #[builder(default = 2)]
        actor_update_interval: usize,
        /// Aggregates the target critics' next-state Q estimates. Canonical
        /// TD3 uses [`SACCriticAggregationMode::Min`].
        #[builder(default = SACCriticAggregationMode::Min)]
        target_aggregation_mode: SACCriticAggregationMode,
        /// Optionally aggregates all online critics for the actor objective.
        /// `None` preserves canonical TD3 behavior by using the first critic.
        actor_aggregation_mode: Option<SACCriticAggregationMode>,
        /// Maximum number of transitions retained in replay.
        #[builder(default = 1_000_000)]
        replay_capacity: usize,
        /// Number of replay transitions sampled per optimization.
        #[builder(default = 256)]
        batch_size: usize,
        /// Collected-transition interval between replay optimizations.
        #[builder(default = 1)]
        update_frequency: usize,
        /// Number of random-action transitions collected before optimization.
        #[builder(default = 1_000)]
        training_start: usize,
        /// Global collected-transition horizon.
        training_horizon: usize,
        #[builder(default = candle_core::DType::F32)] dtype: candle_core::DType,
    ) -> DeterministicActorCriticResult<Self, GE, SE> {
        let strategy = TD3Strategy {
            target_policy_noise,
            target_noise_clip,
            actor_update_interval,
            target_aggregation_mode,
            actor_aggregation_mode,
        };
        let inner = DeterministicActorCriticAgent::builder()
            .online_actor(online_actor)
            .target_actor(target_actor)
            .online_actor_vars(online_actor_vars)
            .target_actor_vars(target_actor_vars)
            .actor_optimizer(actor_optimizer)
            .critics(critics)
            .strategy(strategy)
            .action_space(action_space)
            .observation_space(observation_space)
            .replay_storage_config(replay_storage_config)
            .gamma(gamma)
            .tau(tau)
            .exploration_noise(exploration_noise)
            .replay_capacity(replay_capacity)
            .batch_size(batch_size)
            .update_frequency(update_frequency)
            .training_start(training_start)
            .training_horizon(training_horizon)
            .dtype(dtype)
            .build()?;
        Ok(Self {
            inner,
            logging_info: logger.map(|logger| TD3LoggingInfo { logger }),
        })
    }

    /// Returns the bounded continuous action space used by the agent.
    pub fn get_action_space(&self) -> &BoxSpace {
        self.inner.get_action_space()
    }

    /// Returns the observation-space contract used by the agent.
    pub fn get_observation_space(&self) -> &dyn Space<Error = SE> {
        self.inner.get_observation_space()
    }

    /// Maps observations `[batch, ...observation_shape]` to deterministic
    /// actions `[batch, ...action_shape]`.
    pub fn act_deterministic(
        &self,
        observation: &Tensor,
    ) -> Result<Tensor, DeterministicActorCriticError<GE, SE>> {
        self.inner.act_deterministic(observation)
    }
}

impl<'a, AO, CO, GE, SE, I> Agent<I> for TD3Agent<'a, AO, CO, GE, SE, I>
where
    AO: Optimizer,
    CO: Optimizer,
    GE: Debug,
    SE: Debug,
{
    type Error = DeterministicActorCriticError<GE, SE>;
    type GymError = GE;
    type SpaceError = SE;

    /// Maps observations `[batch, ...observation_shape]` to exploration-noised
    /// actions `[batch, ...action_shape]`.
    fn act(&mut self, observation: &Tensor) -> Result<Tensor, Self::Error> {
        self.inner.act(observation)
    }

    #[cfg_attr(
        feature = "tracing",
        tracing::instrument(
            name = "td3.learn",
            target = "modurl::performance",
            skip_all,
            fields(num_timesteps)
        )
    )]
    fn learn(
        &mut self,
        env: &mut dyn MultiGym<I, Error = GE, SpaceError = SE>,
        num_timesteps: usize,
    ) -> Result<(), Self::Error> {
        self.inner.learn(env, num_timesteps, &mut self.logging_info)
    }
}

#[cfg(test)]
mod tests {
    use super::{TD3Agent, TD3Logger, TD3Strategy};
    use crate::{
        agents::{
            Agent, ReplayDeviceStrategy, ReplayStorageConfig,
            deterministic_actor_critic::{
                DeterministicActorCriticLogEntry, DeterministicActorCriticStrategy,
            },
            sac::{SACCritic, SACCriticAggregationMode, ScalarStateActionCritic},
            test_support::{CountingOptimizer, FixedContinuousEnv},
        },
        gym::{MultiGym, VectorizedGymWrapper},
        models::MLP,
        spaces::BoxSpace,
    };
    use candle_core::{DType, Device, Tensor};
    use candle_nn::{VarBuilder, VarMap};

    #[test]
    fn strategy_uses_the_configured_critic_aggregation() {
        let strategy = TD3Strategy {
            target_policy_noise: 0.2,
            target_noise_clip: 0.5,
            actor_update_interval: 2,
            target_aggregation_mode: SACCriticAggregationMode::Mean,
            actor_aggregation_mode: Some(SACCriticAggregationMode::Mean),
        };
        let values = || {
            vec![
                Tensor::new(&[1.0f32, 3.0], &Device::Cpu).unwrap(),
                Tensor::new(&[3.0f32, 1.0], &Device::Cpu).unwrap(),
                Tensor::new(&[2.0f32, 2.0], &Device::Cpu).unwrap(),
            ]
        };

        assert_eq!(
            strategy
                .aggregate_target_values(values())
                .unwrap()
                .to_vec1::<f32>()
                .unwrap(),
            vec![2.0, 2.0]
        );
        assert_eq!(
            strategy
                .aggregate_actor_values(values())
                .unwrap()
                .to_vec1::<f32>()
                .unwrap(),
            vec![2.0, 2.0]
        );
    }

    #[derive(Default)]
    struct RecordingLogger {
        actor_updates: Vec<bool>,
    }

    impl TD3Logger for RecordingLogger {
        fn log(&mut self, entry: &DeterministicActorCriticLogEntry) {
            self.actor_updates.push(entry.actor_updated);
            assert_eq!(entry.critic_losses.len(), 3);
            assert_eq!(entry.critic_q_values.len(), 3);
            assert!(
                entry
                    .critic_q_values
                    .iter()
                    .all(|values| values.dims() == [1])
            );
            assert_eq!(entry.bellman_targets.dims(), &[1]);
            assert_eq!(entry.replay_actions.dims(), &[1, 1]);
            assert_eq!(entry.critic_learning_rates, vec![1e-3, 1e-3, 1e-3]);
            assert_eq!(entry.actor_loss.is_some(), entry.actor_updated);
            assert_eq!(entry.policy_q_values.is_some(), entry.actor_updated);
            assert_eq!(entry.policy_actions.is_some(), entry.actor_updated);
        }
    }

    fn actor(vars: &VarMap, device: &Device) -> MLP {
        MLP::builder()
            .input_size(4)
            .output_size(1)
            .hidden_layer_sizes(vec![4])
            .output_activation(Tensor::tanh)
            .vb(VarBuilder::from_varmap(vars, DType::F32, device))
            .build()
            .unwrap()
    }

    fn critic<'a>(
        online_vars: &'a VarMap,
        target_vars: &'a mut VarMap,
        device: &Device,
    ) -> SACCritic<'a, CountingOptimizer> {
        let network = |vars| {
            MLP::builder()
                .input_size(5)
                .output_size(1)
                .hidden_layer_sizes(vec![4])
                .vb(VarBuilder::from_varmap(vars, DType::F32, device))
                .build()
                .unwrap()
        };
        SACCritic::builder()
            .online_network(ScalarStateActionCritic::new(network(online_vars)))
            .target_network(ScalarStateActionCritic::new(network(target_vars)))
            .online_vars(online_vars)
            .target_vars(target_vars)
            .optimizer(CountingOptimizer::with_learning_rate(1e-3))
            .build()
            .unwrap()
    }

    #[test]
    fn public_agent_supports_generalized_critic_ensembles_and_delays_actor_updates() {
        let device = Device::Cpu;
        let mut env: VectorizedGymWrapper<FixedContinuousEnv> = vec![
            FixedContinuousEnv::new(device.clone()),
            FixedContinuousEnv::new(device.clone()),
        ]
        .into();
        let online_actor_vars = VarMap::new();
        let mut target_actor_vars = VarMap::new();
        let online_critic_vars_1 = VarMap::new();
        let mut target_critic_vars_1 = VarMap::new();
        let online_critic_vars_2 = VarMap::new();
        let mut target_critic_vars_2 = VarMap::new();
        let online_critic_vars_3 = VarMap::new();
        let mut target_critic_vars_3 = VarMap::new();
        let critic_1 = critic(&online_critic_vars_1, &mut target_critic_vars_1, &device);
        let critic_2 = critic(&online_critic_vars_2, &mut target_critic_vars_2, &device);
        let critic_3 = critic(&online_critic_vars_3, &mut target_critic_vars_3, &device);
        let mut logger = RecordingLogger::default();
        let mut agent = TD3Agent::builder()
            .online_actor(actor(&online_actor_vars, &device))
            .target_actor(actor(&target_actor_vars, &device))
            .online_actor_vars(&online_actor_vars)
            .target_actor_vars(&mut target_actor_vars)
            .actor_optimizer(CountingOptimizer::with_learning_rate(1e-3))
            .critics(vec![critic_1, critic_2, critic_3])
            .action_space(BoxSpace::new_with_universal_bounds(
                vec![1],
                -1.0,
                1.0,
                &device,
            ))
            .observation_space(env.observation_space())
            .replay_storage_config(ReplayStorageConfig::new(ReplayDeviceStrategy::OneDevice(
                device,
            )))
            .exploration_noise(0.0)
            .target_policy_noise(0.0)
            .target_aggregation_mode(SACCriticAggregationMode::Mean)
            .actor_aggregation_mode(SACCriticAggregationMode::Median)
            .replay_capacity(4)
            .batch_size(1)
            .training_start(1)
            .training_horizon(2)
            .logger(&mut logger)
            .build()
            .unwrap();

        agent.learn(&mut env, 2).unwrap();
        assert_eq!(agent.inner.actor_optimizer.steps, 1);
        assert_eq!(agent.inner.critics[0].optimizer().steps, 2);
        assert_eq!(agent.inner.critics[1].optimizer().steps, 2);
        assert_eq!(agent.inner.critics[2].optimizer().steps, 2);
        assert_eq!(logger.actor_updates, vec![false, true]);
    }
}