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

use std::sync::atomic::{AtomicUsize, Ordering};
use std::sync::Arc;

use crate::low_level::{
    Action, ActionMask, ActionSpace, ActionType, CompactFeatures, Entity, Environment, ObsSpace,
    Observation,
};
use crossbeam::channel::{bounded, Receiver, Sender};

use super::{ActionReceiver, Agent, Featurizable, InnerActionReceiver, Obs};

/// An [`Environment`] implementation that is paired with a [`TrainAgent`].
///
/// To create a [`TrainAgent`], use [`TrainEnvBuilder`].
pub struct TrainAgentEnv {
    obs_space: ObsSpace,
    action_space: Vec<(String, ActionSpace)>,
    action: Vec<Sender<u64>>,
    observation: Vec<Receiver<Observation>>,
}

/// Used during training to interface with an external agent implementation.
///
/// To create a [`TrainAgent`], use [`TrainEnvBuilder`].
pub struct TrainAgent {
    action: Receiver<u64>,
    observation: Sender<Observation>,
    entity_names: Vec<String>,
    score: Option<f32>,

    // The `obs_remaining` counters are shared between all agents that connect to the same
    // environment and is used to detect deadlocks that would be caused by an agent awaiting
    // an action before all agents have received an observation.
    // When acting, the agent decrements the counter. When receiving an observation, the agent
    // increments the counter. To prevent an agent from observing a counter reset before it has
    // reached the next step, we use two counters which we swap every step.
    // The `iremaining` variable keeps track of the counter that is currently in use.
    obs_remaining: [Arc<AtomicUsize>; 2],
    iremaining: usize,
    observation_sent: bool,
    agent_count: usize,
}

/// Used to crate a [`TrainAgent`]/[`TrainAgentEnv`] pair which can be used to train an agent.
///
/// The [`TrainAgent`]/[`TrainAgentEnv`] are used to interface with an external Python training training framework.
/// The [`TrainAgent`] is an [`Agent`] implementation that simply forwards all actions to the corresponding
/// [`TrainAgentEnv`], and blocks on the [`TrainAgentEnv`] to receive actions. The [`TrainAgentEnv`] implements
/// the [`Environment`] trait and is used by the Python training framework to single-step the application.
///
/// # Examples
/// Setting up training requires [PyO3](https://pyo3.rs) and a certain amount of boilerplate to define a Python API:
///
/// ```rust
/// use std::sync::Arc;
/// use std::thread;
///
/// use entity_gym_rs::agent::{TrainAgent, TrainAgentEnv, TrainEnvBuilder, Action, Featurizable, Agent};
/// use entity_gym_rs::low_level::py_vec_env::PyVecEnv;
/// use entity_gym_rs::low_level::VecEnv;
/// use pyo3::prelude::*;
///
/// #[derive(Featurizable)]
/// struct Head;
///
/// #[derive(Featurizable)]
/// struct Food;
///
/// #[derive(Action)]
/// enum Move { Up, Down, Left, Right }
///
/// fn run_headless(agent: Box<dyn Agent>, seed: u64) {
///    // Runs the environment.
///    todo!()
/// }
///
/// pub fn spawn_env(seed: u64) -> TrainAgentEnv {
///     // The `TrainEnvBuilder` declares all entity and action types
///     // that will be used by the agent.
///     let (env, agent) = TrainEnvBuilder::default()
///         .entity::<Head>()
///         .entity::<Food>()
///         .action::<Move>()
///         .build();
///     // Start a separate thread to drive execution of the environment.
///     thread::spawn(move || {
///         run_headless(Box::new(agent), seed);
///     });
///     env
/// }
///
/// // Defines a function that will be accessible from Python which creates multiple environments.
/// #[pyfunction]
/// fn create_env(num_envs: usize, threads: usize, first_env_index: u64) -> PyVecEnv {
///     // Boilerplate
///     PyVecEnv {
///         env: VecEnv::new(
///             Arc::new(move |seed| spawn_env(seed)),
///             num_envs,
///             threads,
///             first_env_index,
///         ),
///     }
/// }
///
/// // This defines a Python module and adds our function to it.
/// #[pymodule]
/// fn bevy_snake_enn(_py: Python, m: &PyModule) -> PyResult<()> {
///     m.add_function(wrap_pyfunction!(create_env, m)?)?;
///     Ok(())
/// }
///```
///
/// We now use maturin to build a Python package that exports our environment
/// in a way that is compatible with the Python [enn-trainer](https://github.com/entity-neural-network/enn-trainer) training framework.
/// See [examples/bevy_snake](https://github.com/entity-neural-network/entity-gym-rs/tree/main/examples/bevy_snake) for a full example
/// of how to run training.
#[derive(Debug, Clone, Default)]
pub struct TrainEnvBuilder {
    entities: Vec<(String, Entity)>,
    actions: Vec<(String, ActionSpace)>,
}

impl Environment for TrainAgentEnv {
    fn obs_space(&self) -> ObsSpace {
        self.obs_space.clone()
    }

    fn action_space(&self) -> Vec<(ActionType, ActionSpace)> {
        self.action_space.clone()
    }

    fn agents(&self) -> usize {
        self.action.len()
    }

    fn reset(&mut self) -> Vec<Box<Observation>> {
        self.observation
            .iter()
            .map(|obs| Box::new(obs.recv().unwrap()))
            .collect()
    }

    fn act(&mut self, action: &[Vec<Option<Action>>]) -> Vec<Box<Observation>> {
        assert!(action.len() == self.action.len());
        for (sender, action) in self.action.iter().zip(action.iter()) {
            assert!(action.len() == 1);
            match &action[0] {
                Some(Action::Categorical { actors: _, action }) => {
                    assert!(action.len() == 1);
                    sender.send(action[0] as u64).unwrap();
                }
                Some(_) => panic!("unexpected action"),
                None => {
                    panic!("No action provided");
                }
            }
        }
        self.observation
            .iter()
            .map(|obs| Box::new(obs.recv().unwrap()))
            .collect()
    }
}

impl Agent for TrainAgent {
    fn act_dyn(&mut self, action: &str, _num_actions: u64, obs: &Obs) -> Option<u64> {
        self.send_obs_raw(action, obs);
        let remaining = self.obs_remaining[self.iremaining].load(Ordering::SeqCst);
        if remaining != 0 && (remaining != self.agent_count || self.agent_count == 1) {
            panic!("TrainAgent::act called before all agents have received observations. This is not allowed. If you have multiple agents, call the `act_async` on every agent before awaiting any actions.");
        }
        match self.action.recv() {
            Ok(action) => {
                self.obs_remaining[self.iremaining].store(self.agent_count, Ordering::SeqCst);
                self.iremaining = 1 - self.iremaining;
                self.observation_sent = false;
                Some(action)
            }
            Err(_) => None,
        }
    }

    fn act_async_dyn(&mut self, action: &str, _num_actions: u64, obs: &Obs) -> ActionReceiver<u64> {
        self.send_obs_raw(action, obs);
        self.observation_sent = false;
        self.iremaining = 1 - self.iremaining;
        ActionReceiver {
            inner: InnerActionReceiver::Receiver {
                receiver: self.action.clone(),
                observations_remaining: self.obs_remaining[1 - self.iremaining].clone(),
                agent_count: self.agent_count,
                phantom: Default::default(),
            },
        }
    }

    fn game_over(&mut self, obs: &Obs) {
        let obs = Observation {
            features: CompactFeatures {
                counts: vec![0; self.entity_names.len()],
                data: vec![],
            },
            ids: vec![None],
            actions: vec![None],
            done: true,
            reward: obs.score - self.score.unwrap_or(0.0),
            metrics: obs.metrics.clone(),
        };
        self.score = None;
        let _ = self.observation.send(obs);
    }
}

impl TrainAgent {
    fn send_obs_raw(&mut self, _action: &str, obs: &Obs) {
        assert!(
            self.observation_sent == false,
            "Observation already sent, await the next action before sending a new observation."
        );
        self.obs_remaining[self.iremaining].fetch_sub(1, Ordering::SeqCst);

        let mut data = vec![];
        let mut counts = vec![];
        for name in &self.entity_names {
            match obs.entities.get(name.as_str()) {
                Some((feats, count, _)) => {
                    data.extend(feats.iter());
                    counts.push(*count);
                }
                None => {
                    counts.push(0);
                }
            }
        }
        // TODO: make noise when obs contains entity that is not in obs space
        let last_score = self.score.replace(obs.score).unwrap_or(obs.score);
        let observation = Observation {
            features: CompactFeatures { counts, data },
            ids: vec![None],
            actions: vec![Some(ActionMask::DenseCategorical {
                actors: vec![0],
                mask: None,
            })],
            done: obs.done,
            reward: obs.score - last_score,
            metrics: obs.metrics.clone(),
        };
        let _ = self.observation.send(observation);
    }
}

impl TrainEnvBuilder {
    /// Registers the type of an observable entity.
    pub fn entity<E: Featurizable>(mut self) -> Self {
        assert!(
            self.entities.iter().all(|(n, _)| n != E::name()),
            "Already have an entity with name \"{}\"",
            E::name(),
        );
        self.entities.push((
            E::name().to_string(),
            Entity {
                features: E::feature_names().iter().map(|n| n.to_string()).collect(),
            },
        ));
        self
    }

    /// Registers the type of an action.
    pub fn action<'a, A: super::Action<'a>>(mut self) -> Self {
        assert!(
            self.actions.iter().all(|(n, _)| n != A::name()),
            "Already have an action with name \"{}\"",
            A::name(),
        );
        self.actions.push((
            A::name().to_string(),
            ActionSpace::Categorical {
                choices: A::labels().iter().map(|c| c.to_string()).collect(),
            },
        ));
        self
    }

    /// Creates a new [`TrainAgentEnv`] which is connected to multiple [`TrainAgent`]s.
    /// When using multiple agents in an environment, it is required to call `act_async`
    /// on each agent before awaiting any actions.
    pub fn build_multiagent(self, num_agents: usize) -> (TrainAgentEnv, Vec<TrainAgent>) {
        let mut environment = TrainAgentEnv {
            obs_space: ObsSpace {
                entities: self.entities.to_vec(),
            },
            action_space: self.actions,
            action: vec![],
            observation: vec![],
        };
        let mut agents = vec![];
        let obs_remaining = [
            Arc::new(AtomicUsize::new(num_agents)),
            Arc::new(AtomicUsize::new(num_agents)),
        ];
        for _ in 0..num_agents {
            let (action_tx, action_rx) = bounded(1);
            let (observation_tx, observation_rx) = bounded(1);
            let entity_names = self.entities.iter().map(|(n, _)| n.to_string()).collect();
            environment.action.push(action_tx);
            environment.observation.push(observation_rx);
            agents.push(TrainAgent {
                action: action_rx,
                observation: observation_tx,
                entity_names,
                score: None,
                obs_remaining: obs_remaining.clone(),
                iremaining: 0,
                observation_sent: false,
                agent_count: num_agents,
            });
        }

        (environment, agents)
    }

    /// Creates a new [`TrainAgentEnv`]/[`TrainAgent`] pair.
    pub fn build(self) -> (TrainAgentEnv, TrainAgent) {
        let (action_tx, action_rx) = bounded(1);
        let (observation_tx, observation_rx) = bounded(1);
        let entity_names = self.entities.iter().map(|(n, _)| n.to_string()).collect();
        (
            TrainAgentEnv {
                obs_space: ObsSpace {
                    entities: self.entities.into_iter().collect(),
                },
                action_space: self.actions,
                action: vec![action_tx],
                observation: vec![observation_rx],
            },
            TrainAgent {
                action: action_rx,
                observation: observation_tx,
                entity_names,
                score: None,
                obs_remaining: [Arc::new(AtomicUsize::new(1)), Arc::new(AtomicUsize::new(1))],
                iremaining: 0,
                observation_sent: false,
                agent_count: 1,
            },
        )
    }
}