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

use crate::history::Features;
use crate::history::History;
use crate::network::Network;
use crate::node::Node;
use crate::settings::Settings;
use crate::training_sample_generator::TrainingSampleGenerator;

#[derive(Debug)]
pub struct Trainer<'a> {
    settings: &'a Settings,
    history: &'a mut History,
}

impl<'a> Trainer<'a> {
    pub fn new(settings: &'a Settings, history: &'a mut History) -> Trainer<'a> {
        Trainer { settings, history }
    }

    pub fn train(&mut self) {
        let lr = 0.000_005;
        let momentum = 0.0;
        let batch_size = 1000;
        let plateau_threshold = 3000;
        let generator = TrainingSampleGenerator::new(self.settings, self.history);

        println!(
            "Evaluating error rate on current {:#?}",
            self.history.network
        );
        let mut best_overall_network = self.history.network;
        let mut best_overall_error = self
            .history
            .network
            .average_error(&generator, batch_size * 10);
        println!("Current network error rate is {}", best_overall_error);

        loop {
            let mut best_restart_network = Network::random();
            let mut best_restart_error = 10000.0;
            let mut cycles_since_best_restart_error = 0;
            let mut network = Network::random();
            let mut node_increments = [Node::default(), Node::default(), Node::default()];
            let mut output_increments = [0.0, 0.0, 0.0, 0.0];

            loop {
                let mut batch_error = 0.0;
                let mut batch_samples = 0.0;

                // Two node network example:
                // (Note: we currently are using a three node version, s_0 to s_2 and o_0 to o_2.)
                //
                //             b_1
                //                \
                //        f_1 --- s_1 -- o_1
                //            \ /           \
                //             x       b_3 -- s_3 -> o_3 -> e
                //            / \           /
                //        f_2 --- s_2 -- o_2
                //                /
                //             b_2
                //
                // Error (e) = 0.5(t - o_3)^2
                // Final output (o_3) = tanh(s_3)
                // Final sum (s_3) = b_3 + w3_1*o_1 + w3_2*o_2
                // Hidden node 1 output (o_1) = tanh(s_1)
                // Hidden node 1 sum (s_1) = b_1 + w1_1*f_1 + w1_2*f_2
                // Hidden node 2 output (o_2) = tanh(s_2)
                // Hidden node 2 sum (s_2) = b_2 + w2_1*f_1 + w2_2*f_2
                // Derivative of error with respect to o_3 (d_e/d_o_3 0.5(t - o_3)^2): -(t - o_3)
                // Derivative of o_3 respect to s_3 (d_o_3/d_s_3 tanh(s_3)): 1.0 - tanh(s_3)^2
                // Derivative of s_3 with respect to weight w3_1 (d_s_3/d_w3_1 bias + w3_1*o_1 + w3_2*o_2): o_1
                // Derivative of error with respect to weight w3_1 (d_e/d_o_3 * d_o_3/d_s_3 * d_s_3/d_w3_1): -(t - o_3) * (1 - tanh(s_3)^2) * o_1
                // Derivative of s_3 with respect to o_1 (d_s_3/d_o_1 b_3 + w3_1*o_1 + w3_2*o_2): w3_1
                // Derivative of o_1 with respect to s_1 (d_o_1/d_s_1): 1.0 - tanh(s_1)^2
                // Derivative of s_1 with respect to weight w1_1 (d_s_1/d_w1_1): f_1
                // Full derivation of o_3: tanh(b_3 + w3_1*tanh(b_1 + w1_1*f_1 + w1_2*f_2) + w3_2*tanh(b_2 + w2_1*f_1 + w2_2*f_2))
                // Full derivation of s_3: b_3 + w3_1*tanh(b_1 + w1_1*f_1 + w1_2*f_2) + w3_2*tanh(b_2 + w2_1*f_1 + w2_2*f_2)
                // Full error derivation: 0.5(t - tanh(b_3 + w3_1*tanh(b_1 + w1_1*f_1 + w1_2*f_2) + w3_2*tanh(b_2 + w2_1*f_1 + w2_2*f_2)))^2
                // Full derivative for o_1: -(t - o_3) * (1.0 - tanh(s_3)^2) * w3_1
                // Full derivative for w1_1: -(t - o_3) * (1.0 - tanh(s_3)^2) * w3_1 * (1.0 - tanh(s_1)^2) * f_1
                // Checked: https://www.wolframcloud.com/objects/617707c2-5016-4fb3-b73d-bd688b884967
                generator.generate(Some(batch_size), |features: &Features, correct: bool| {
                    let target = if correct { 1.0 } else { -1.0 };
                    network.compute(features);

                    let error = 0.5 * (target - network.final_output).powi(2);
                    batch_error += error;
                    batch_samples += 1.0;

                    let d_e_d_o_3 = -(target - network.final_output);
                    let d_o_3_d_s_3 = 1.0 - network.final_sum.tanh().powi(2);

                    // Output bias
                    output_increments[0] =
                        momentum * output_increments[0] + lr * d_e_d_o_3 * d_o_3_d_s_3 * 1.0;
                    // Final sum node 1 output weight
                    output_increments[1] = momentum * output_increments[1]
                        + lr * d_e_d_o_3 * d_o_3_d_s_3 * network.hidden_node_outputs[0];
                    // Final sum node 2 output weight
                    output_increments[2] = momentum * output_increments[2]
                        + lr * d_e_d_o_3 * d_o_3_d_s_3 * network.hidden_node_outputs[1];
                    // Final sum node 3 output weight
                    output_increments[3] = momentum * output_increments[3]
                        + lr * d_e_d_o_3 * d_o_3_d_s_3 * network.hidden_node_outputs[2];

                    let d_s_3_d_o_0 = network.final_weights[0];
                    let d_s_3_d_o_1 = network.final_weights[1];
                    let d_s_3_d_o_2 = network.final_weights[2];
                    let d_o_0_d_s_0 = 1.0 - network.hidden_node_sums[0].tanh().powi(2);
                    let d_o_1_d_s_1 = 1.0 - network.hidden_node_sums[1].tanh().powi(2);
                    let d_o_2_d_s_2 = 1.0 - network.hidden_node_sums[2].tanh().powi(2);
                    let d_e_d_s_0 = d_e_d_o_3 * d_o_3_d_s_3 * d_s_3_d_o_0 * d_o_0_d_s_0;
                    let d_e_d_s_1 = d_e_d_o_3 * d_o_3_d_s_3 * d_s_3_d_o_1 * d_o_1_d_s_1;
                    let d_e_d_s_2 = d_e_d_o_3 * d_o_3_d_s_3 * d_s_3_d_o_2 * d_o_2_d_s_2;

                    node_increments[0].offset =
                        momentum * node_increments[0].offset + lr * d_e_d_s_0 * 1.0;
                    node_increments[0].age =
                        momentum * node_increments[0].age + lr * d_e_d_s_0 * features.age_factor;
                    node_increments[0].length = momentum * node_increments[0].length
                        + lr * d_e_d_s_0 * features.length_factor;
                    node_increments[0].exit =
                        momentum * node_increments[0].exit + lr * d_e_d_s_0 * features.exit_factor;
                    node_increments[0].recent_failure = momentum
                        * node_increments[0].recent_failure
                        + lr * d_e_d_s_0 * features.recent_failure_factor;
                    node_increments[0].selected_dir = momentum * node_increments[0].selected_dir
                        + lr * d_e_d_s_0 * features.selected_dir_factor;
                    node_increments[0].dir =
                        momentum * node_increments[0].dir + lr * d_e_d_s_0 * features.dir_factor;
                    node_increments[0].overlap = momentum * node_increments[0].overlap
                        + lr * d_e_d_s_0 * features.overlap_factor;
                    node_increments[0].immediate_overlap = momentum
                        * node_increments[0].immediate_overlap
                        + lr * d_e_d_s_0 * features.immediate_overlap_factor;
                    node_increments[0].selected_occurrences = momentum
                        * node_increments[0].selected_occurrences
                        + lr * d_e_d_s_0 * features.selected_occurrences_factor;
                    node_increments[0].occurrences = momentum * node_increments[0].occurrences
                        + lr * d_e_d_s_0 * features.occurrences_factor;

                    node_increments[1].offset =
                        momentum * node_increments[1].offset + lr * d_e_d_s_1 * 1.0;
                    node_increments[1].age =
                        momentum * node_increments[1].age + lr * d_e_d_s_1 * features.age_factor;
                    node_increments[1].length = momentum * node_increments[1].length
                        + lr * d_e_d_s_1 * features.length_factor;
                    node_increments[1].exit =
                        momentum * node_increments[1].exit + lr * d_e_d_s_1 * features.exit_factor;
                    node_increments[1].recent_failure = momentum
                        * node_increments[1].recent_failure
                        + lr * d_e_d_s_1 * features.recent_failure_factor;
                    node_increments[1].selected_dir = momentum * node_increments[1].selected_dir
                        + lr * d_e_d_s_1 * features.selected_dir_factor;
                    node_increments[1].dir =
                        momentum * node_increments[1].dir + lr * d_e_d_s_1 * features.dir_factor;
                    node_increments[1].overlap = momentum * node_increments[1].overlap
                        + lr * d_e_d_s_1 * features.overlap_factor;
                    node_increments[1].immediate_overlap = momentum
                        * node_increments[1].immediate_overlap
                        + lr * d_e_d_s_1 * features.immediate_overlap_factor;
                    node_increments[1].selected_occurrences = momentum
                        * node_increments[1].selected_occurrences
                        + lr * d_e_d_s_1 * features.selected_occurrences_factor;
                    node_increments[1].occurrences = momentum * node_increments[1].occurrences
                        + lr * d_e_d_s_1 * features.occurrences_factor;

                    node_increments[2].offset =
                        momentum * node_increments[2].offset + lr * d_e_d_s_2 * 1.0;
                    node_increments[2].age =
                        momentum * node_increments[2].age + lr * d_e_d_s_2 * features.age_factor;
                    node_increments[2].length = momentum * node_increments[2].length
                        + lr * d_e_d_s_2 * features.length_factor;
                    node_increments[2].exit =
                        momentum * node_increments[2].exit + lr * d_e_d_s_2 * features.exit_factor;
                    node_increments[2].recent_failure = momentum
                        * node_increments[2].recent_failure
                        + lr * d_e_d_s_2 * features.recent_failure_factor;
                    node_increments[2].selected_dir = momentum * node_increments[2].selected_dir
                        + lr * d_e_d_s_2 * features.selected_dir_factor;
                    node_increments[2].dir =
                        momentum * node_increments[2].dir + lr * d_e_d_s_2 * features.dir_factor;
                    node_increments[2].overlap = momentum * node_increments[2].overlap
                        + lr * d_e_d_s_2 * features.overlap_factor;
                    node_increments[2].immediate_overlap = momentum
                        * node_increments[2].immediate_overlap
                        + lr * d_e_d_s_2 * features.immediate_overlap_factor;
                    node_increments[2].selected_occurrences = momentum
                        * node_increments[2].selected_occurrences
                        + lr * d_e_d_s_2 * features.selected_occurrences_factor;
                    node_increments[2].occurrences = momentum * node_increments[2].occurrences
                        + lr * d_e_d_s_2 * features.occurrences_factor;

                    let node0 = network.hidden_nodes[0];
                    let node1 = network.hidden_nodes[1];
                    let node2 = network.hidden_nodes[2];
                    network = Network {
                        hidden_nodes: [
                            Node {
                                offset: node0.offset - node_increments[0].offset,
                                age: node0.age - node_increments[0].age,
                                length: node0.length - node_increments[0].length,
                                exit: node0.exit - node_increments[0].exit,
                                recent_failure: node0.recent_failure
                                    - node_increments[0].recent_failure,
                                selected_dir: node0.selected_dir - node_increments[0].selected_dir,
                                dir: node0.dir - node_increments[0].dir,
                                overlap: node0.overlap - node_increments[0].overlap,
                                immediate_overlap: node0.immediate_overlap
                                    - node_increments[0].immediate_overlap,
                                selected_occurrences: node0.selected_occurrences
                                    - node_increments[0].selected_occurrences,
                                occurrences: node0.occurrences - node_increments[0].occurrences,
                            },
                            Node {
                                offset: node1.offset - node_increments[1].offset,
                                age: node1.age - node_increments[1].age,
                                length: node1.length - node_increments[1].length,
                                exit: node1.exit - node_increments[1].exit,
                                recent_failure: node1.recent_failure
                                    - node_increments[1].recent_failure,
                                selected_dir: node1.selected_dir - node_increments[1].selected_dir,
                                dir: node1.dir - node_increments[1].dir,
                                overlap: node1.overlap - node_increments[1].overlap,
                                immediate_overlap: node1.immediate_overlap
                                    - node_increments[1].immediate_overlap,
                                selected_occurrences: node1.selected_occurrences
                                    - node_increments[1].selected_occurrences,
                                occurrences: node1.occurrences - node_increments[1].occurrences,
                            },
                            Node {
                                offset: node2.offset - node_increments[2].offset,
                                age: node2.age - node_increments[2].age,
                                length: node2.length - node_increments[2].length,
                                exit: node2.exit - node_increments[2].exit,
                                recent_failure: node2.recent_failure
                                    - node_increments[2].recent_failure,
                                selected_dir: node2.selected_dir - node_increments[2].selected_dir,
                                dir: node2.dir - node_increments[2].dir,
                                overlap: node2.overlap - node_increments[2].overlap,
                                immediate_overlap: node2.immediate_overlap
                                    - node_increments[2].immediate_overlap,
                                selected_occurrences: node2.selected_occurrences
                                    - node_increments[2].selected_occurrences,
                                occurrences: node2.occurrences - node_increments[2].occurrences,
                            },
                        ],
                        hidden_node_sums: [0.0, 0.0, 0.0],
                        hidden_node_outputs: [0.0, 0.0, 0.0],
                        final_bias: network.final_bias - output_increments[0],
                        final_weights: [
                            network.final_weights[0] - output_increments[1],
                            network.final_weights[1] - output_increments[2],
                            network.final_weights[2] - output_increments[3],
                        ],
                        final_sum: 0.0,
                        final_output: 0.0,
                    };
                });

                if batch_error / batch_samples < best_restart_error {
                    best_restart_error = batch_error / batch_samples;
                    best_restart_network = network;
                    cycles_since_best_restart_error = 0;
                } else {
                    cycles_since_best_restart_error += 1;
                    if cycles_since_best_restart_error > plateau_threshold {
                        println!("Plateaued at {}.", batch_error / batch_samples);

                        if best_restart_error < best_overall_error {
                            best_overall_error = best_restart_error;
                            best_overall_network = best_restart_network;

                            println!(
                                "New best overall for {:#?} with error {} (new best)",
                                best_overall_network, best_overall_error
                            );
                        } else {
                            println!(
                                "Best overall remains {:#?} with error {} (old)",
                                best_overall_network, best_overall_error
                            );
                        }
                        break;
                    }
                }

                //                println!("Error of {} (vs {} {} ago)", batch_error / batch_samples, best_restart_error, cycles_since_best_restart_error);
            }
        }
    }
}