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
use core::{Algorithm, Controller, Predictor, Shared, Parameter, Vector, Matrix, Trace};
use domains::Transition;
use fa::{Approximator, Parameterised, MultiLFA, Projection, Projector, QFunction};
use policies::{fixed::Greedy, Policy};
use std::marker::PhantomData;
pub struct QLambda<S, M: Projector<S>, P: Policy<S>> {
trace: Trace,
pub fa_theta: Shared<MultiLFA<S, M>>,
pub policy: Shared<P>,
pub target: Greedy<S>,
pub alpha: Parameter,
pub gamma: Parameter,
phantom: PhantomData<S>,
}
impl<S: 'static, M: Projector<S> + 'static, P: Policy<S>> QLambda<S, M, P> {
pub fn new<T1, T2>(
trace: Trace,
fa_theta: Shared<MultiLFA<S, M>>,
policy: Shared<P>,
alpha: T1,
gamma: T2,
) -> Self
where
T1: Into<Parameter>,
T2: Into<Parameter>,
{
QLambda {
trace,
fa_theta: fa_theta.clone(),
policy,
target: Greedy::new(fa_theta),
alpha: alpha.into(),
gamma: gamma.into(),
phantom: PhantomData,
}
}
}
impl<S, M: Projector<S>, P: Policy<S, Action = usize>> QLambda<S, M, P> {
#[inline(always)]
fn update_theta(&mut self, action: P::Action, error: f64) {
self.fa_theta.borrow_mut().update_action_phi(
&Projection::Dense(self.trace.get()),
action,
self.alpha * error,
);
}
#[inline(always)]
fn update_trace(&mut self, state: &S, action: P::Action, phi_s: Projection) {
if action == self.target.sample(state) {
let rate = self.trace.lambda.value() * self.gamma.value();
self.trace.decay(rate);
} else {
self.trace.decay(0.0);
}
self.trace.update(&phi_s.expanded(self.fa_theta.borrow().projector.dim()));
}
}
impl<S, M: Projector<S>, P: Policy<S, Action = usize>> Algorithm<S, P::Action> for QLambda<S, M, P> {
fn handle_sample(&mut self, t: &Transition<S, P::Action>) {
let (s, ns) = (t.from.state(), t.to.state());
let phi_s = self.fa_theta.borrow().projector.project(s);
let na = self.target.sample(&ns);
let qsa = self.fa_theta.borrow().evaluate_action_phi(&phi_s, t.action);
let nqsna = self.fa_theta.borrow().evaluate_action(ns, na);
let td_error = t.reward + self.gamma * nqsna - qsa;
self.update_trace(s, t.action, phi_s);
self.update_theta(t.action, td_error);
}
fn handle_terminal(&mut self, t: &Transition<S, P::Action>) {
{
let s = t.from.state();
let phi_s = self.fa_theta.borrow().projector.project(s);
let qsa = self.fa_theta.borrow().evaluate_action_phi(&phi_s, t.action);
self.update_trace(s, t.action, phi_s);
self.update_theta(t.action, t.reward - qsa);
self.target.handle_terminal(t);
self.policy.borrow_mut().handle_terminal(t);
self.trace.decay(0.0);
}
self.alpha = self.alpha.step();
self.gamma = self.gamma.step();
}
}
impl<S, M: Projector<S>, P: Policy<S, Action = usize>> Controller<S, P::Action> for QLambda<S, M, P> {
fn sample_target(&mut self, s: &S) -> P::Action { self.target.sample(s) }
fn sample_behaviour(&mut self, s: &S) -> P::Action { self.policy.borrow_mut().sample(s) }
}
impl<S, M: Projector<S>, P: Policy<S, Action = usize>> Predictor<S, P::Action> for QLambda<S, M, P> {
fn predict_v(&mut self, s: &S) -> f64 {
let a = self.sample_target(s);
self.predict_qsa(s, a)
}
fn predict_qs(&mut self, s: &S) -> Vector<f64> {
self.fa_theta.borrow().evaluate(s).unwrap()
}
fn predict_qsa(&mut self, s: &S, a: P::Action) -> f64 {
self.fa_theta.borrow().evaluate_action(&s, a)
}
}
impl<S, M: Projector<S>, P: Policy<S, Action = usize>> Parameterised for QLambda<S, M, P> {
fn weights(&self) -> Matrix<f64> {
self.fa_theta.borrow().weights()
}
}