beet_ml 0.0.8

Machine Learning actions built upon beet_flow
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

use crate::prelude::*;
use beet_core::prelude::*;
use beet_flow::prelude::*;
use std::marker::PhantomData;

#[action(step_environment::<S>)]
#[derive(Debug, Clone, PartialEq, Component, Reflect)]
#[reflect(Component)]
pub struct StepEnvironment<S: RlSessionTypes>
where
	S::State: Component,
	S::Action: Component,
	S::QLearnPolicy: Component,
	S::Env: Component,
{
	episode: u32,
	step: u32,
	#[reflect(ignore)]
	phantom: PhantomData<S>,
}

impl<S: RlSessionTypes> StepEnvironment<S>
where
	S::State: Component,
	S::Action: Component,
	S::QLearnPolicy: Component,
	S::Env: Component,
{
	pub fn new(episode: u32) -> Self {
		Self {
			episode,
			step: 0,
			phantom: PhantomData,
		}
	}
}


fn step_environment<S: RlSessionTypes>(
	ev: On<GetOutcome>,
	mut rng: ResMut<RandomSource>,
	mut end_episode_events: MessageWriter<EndEpisode<S::EpisodeParams>>,
	mut commands: Commands,
	mut sessions: Query<&mut S::QLearnPolicy>,
	mut agents: AgentQuery<(
		&S::State,
		&mut S::Action,
		&mut S::Env,
		&QLearnParams,
		&SessionEntity,
	)>,
	mut query: Query<&mut StepEnvironment<S>>,
) -> Result
where
	S::State: Component,
	S::Action: Component,
	S::QLearnPolicy: Component,
	S::Env: Component,
{
	let mut step = query.get_mut(ev.action())?;
	let (state, mut action, mut env, params, session_entity) =
		agents.get_mut(ev.action())?;
	let mut table = sessions.get_mut(**session_entity)?;

	let outcome = env.step(&state, &action);
	// we ignore the state of the outcome, allow simulation to determine
	let epsilon = params.epsilon(step.episode);

	*action = table.step(
		params,
		&mut **rng,
		epsilon,
		&action,
		state,
		&outcome.state,
		outcome.reward,
	);
	// log::info!(
	// 	"step complete - action: {:?}, reward: {:?}",
	// 	action,
	// 	outcome.reward
	// );
	commands.entity(ev.action()).trigger_target(Outcome::Pass);
	step.step += 1;

	if outcome.done || step.step >= params.max_steps {
		end_episode_events.write(EndEpisode::new(**session_entity));
	}
	Ok(())
}

#[cfg(test)]
mod test {
	use crate::prelude::*;
	use beet_core::prelude::*;
	use beet_flow::prelude::*;

	#[test]
	fn works() {
		let mut app = App::new();

		let on_result = observer_ext::observe_triggers::<Outcome>(app.world_mut());

		app.add_plugins((
			ControlFlowPlugin::default(),
			RlSessionPlugin::<FrozenLakeEpParams>::default(),
		))
		.init_resource::<RandomSource>()
		.insert_time();

		let map = FrozenLakeMap::default_four_by_four();

		let mut rng = RandomSource::from_seed(0);

		let session = app.world_mut().spawn(FrozenLakeQTable::default()).id();

		app.world_mut()
			.spawn((
				RlAgentBundle {
					state: map.agent_position(),
					action: GridDirection::sample(&mut *rng),
					env: QTableEnv::new(map.transition_outcomes()),
					params: QLearnParams::default(),
					session: SessionEntity(session),
					despawn: DespawnOnEpisodeEnd,
				},
				StepEnvironment::<FrozenLakeQTableSession>::new(0),
			))
			.trigger_target(GetOutcome)
			.flush();


		app.world_mut().insert_resource(rng);

		app.update();

		(on_result.len() > 0).xpect_true();


		let table = app.world().get::<FrozenLakeQTable>(session).unwrap();
		table
			.keys()
			.next()
			.xpect_eq(Some(&GridPos(UVec2::new(0, 0))));
		let inner = table.values().next().unwrap();
		inner.iter().next().unwrap().1.xpect_eq(0.);
	}
}