-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
811 lines (743 loc) · 36.7 KB
/
index.html
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
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<title>Evolutionary Methods for Interpretable Control</title>
<link rel="stylesheet" href="reveal/css/reset.css">
<link rel="stylesheet" href="reveal/css/reveal.css">
<link rel="stylesheet" href="reveal/css/theme/white.css">
<link rel="stylesheet" href="reveal/css/custom.css">
<link rel="stylesheet" href="https://d9w.github.io/interpretable_control/reveal/css/custom.css">
<!-- Theme used for syntax highlighting of code -->
<link rel="stylesheet" href="reveal/lib/css/zenburn.css">
<!-- Printing and PDF exports -->
<script>
var link = document.createElement('link');
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = window.location.search.match(/print-pdf/gi) ? 'reveal/css/print/pdf.css' : 'reveal/css/print/paper.css';
document.getElementsByTagName('head')[0].appendChild(link);
</script>
</head>
<body>
<div class="reveal">
<div class="slides">
<section>
<br />
<br />
<h2>Evolutionary Methods</h2>
<h2>for Interpretable Control</h2>
<br />
<h4>Dennis G. Wilson</h4>
<h4>EvoRL Workshop</h4>
<h4>GECCO 2023</h4>
<br />
</section>
<section>
<section>
<h2>Policy search in critical applications</h2>
<img src="img/l2f_taxi_path.png" width="40%">
<img src="img/rl-taxi2.png" width="40%">
<div class="source">
Wilson D, et al. "Learning Robust and Readable Control Laws for Aircraft Taxi Control." Under
review.
</div>
</section>
<section>
<h2>Autonomous Vehicles</h2>
<br />
<img src="img/self_driving_1.png" width="70%">
<div class="source">
Atakishiyev, Shahin, et al. "Explainable artificial intelligence for autonomous driving: A
comprehensive overview and field guide for future research directions." arXiv preprint
arXiv:2112.11561 (2021).
</div>
</section>
<section>
<h2>How to train your policy</h2>
<br />
<div class="flex-container">
<div class="flex-child">
<img src="img/RL_loop2.png" width="90%">
</div>
<div class="flex-child">
Representation:
<br />
Neural network
<br />
<br />
Optimization:
<br />
<ul>
<li>Value function approximation</li>
<li>Policy gradient</li>
<li>Direct policy search</li>
</ul>
<br />
<br />
Objective:
<br />
\[\begin{aligned}
f(\pi_\theta) = \sum_t^T \gamma^t R(s_t, \pi_\theta(s_t))
\end{aligned} \]
<br />
</div>
</div>
<br />
<div class="source">
Mao, Hongzi, et al. "Resource management with deep reinforcement learning." Proceedings of the
15th ACM workshop on hot topics in networks. 2016.
</div>
</section>
<section>
<h2>Explaining Neural Networks</h2>
<img src="img/self_driving_2.png" width="80%">
<div class="source">
Kim, Jinkyu, and John Canny. "Interpretable learning for self-driving cars by visualizing causal
attention." Proceedings of the IEEE international conference on computer vision. 2017.
</div>
</section>
<section>
<h2>The problem with explaining</h2>
<img src="img/saliency.png" width="80%">
<br />
Attention maps and feature importance analysis explain <b>what</b> information is being used, not
<b>how</b>.
<br />
<br />
<div class="source">
Rudin, Cynthia. "Stop explaining black box machine learning models for high stakes decisions and
use interpretable models instead." Nature machine intelligence 1.5 (2019): 206-215.
</div>
</section>
<section>
<h2>The many forms of explanation</h2>
<img src="img/explaining.png" width="40%">
<br />
Evolution can optimize control policies which are <b>interpretable by design</b>.
<br />
<br />
<div class="source">
Zhou, Ryan, and Ting Hu. "Evolutionary approaches to explainable machine learning." arXiv
preprint arXiv:2306.14786 (2023).
</div>
</section>
</section>
<section>
<section>
<h2>How to get interpretable policies</h2>
<img src="img/imitation_learning.jpg" width="70%">
<br />
<br />
Start with a good neural network policy and <b>imitate it</b>.
<br />
Predict expert action as a <b>classification problem</b>.
<br />
<div class="source">
Delgado, Juan, et al. "Robotics in construction: A critical review
of the reinforcement learning and imitation learning paradigms." Advanced Engineering
Informatics 54 (2022): 101787.
</div>
</section>
<section>
<h2>Imitation with Decision Trees</h2>
<img src="img/viper.png">
<br />Pong inputs extracted from raw images:
<br />ball position $(x, y)$ and velocity $(v_x , v_y)$
<br />player paddle position $y_p$ , velocity $v_p$, acceleration $a_p$, and jerk $j_p$.
<div class="source">
Bastani, Osbert, Yewen Pu, and Armando Solar-Lezama. "Verifiable reinforcement learning via
policy extraction." Advances in neural information processing systems 31 (2018).
</div>
</section>
<section>
<h2>Imitation with Decision Trees</h2>
<img src="img/sdt.png" width="55%">
<br />
Application to 10x10 visual area around Mario.
<div class="source">
Coppens, Youri, et al. "Distilling deep reinforcement learning policies in soft decision trees."
Proceedings of the IJCAI 2019 workshop on explainable artificial intelligence. 2019.
</div>
</section>
<section>
<h2>Imitation with Functional Trees</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/gprl1.jpg" width="60%">
</div>
<div class="flex-child">
<img src="img/gprl2.jpg" width="90%">
</div>
</div>
<br />
Model-based and model-free <b>symbolic regression</b> (genetic programming)
<div class="source">
Hein, Daniel, et al. "Interpretable policies for reinforcement
learning by genetic programming." Engineering Applications of Artificial Intelligence 76 (2018):
158-169.
</div>
</section>
<section>
<h2>Imitation with Programs</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/pirl1.png" width="60%">
</div>
<div class="flex-child">
<img src="img/pirl2.png" width="50%">
</div>
</div>
Example policy:
<br />
<img src="img/pirl3.png" width="80%">
<br />
Similar to <b>grammatical evolution</b> with mutation function "neighborhood_pool"
<div class="source">
Verma, Abhinav, et al. "Programmatically interpretable reinforcement learning." International
Conference on Machine Learning. PMLR, 2018.
</div>
</section>
<section>
<h2>Generating Policy Programs</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/dsp1.png" width="50%">
</div>
<div class="flex-child">
<img src="img/dsp2.png" width="70%">
</div>
</div>
Generate programs with a <b>recurrent network</b> guided by an anchor policy.
<div class="source">
Landajuela, Mikel, et al. "Discovering symbolic policies with deep reinforcement learning."
International Conference on Machine Learning. PMLR, 2021.
</div>
</section>
<section>
<h2>Generation > Imitation</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/dsp3.png" width="70%">
</div>
<div class="flex-child">
<img src="img/dsp4.png" width="60%">
</div>
</div>
Imitation policies may not align with expert policies over full trajectories.
<br />
Generated policies, boosted by an expert policy, outperform imitation.
<div class="source">
Landajuela, Mikel, et al. "Discovering symbolic policies with deep reinforcement learning."
International Conference on Machine Learning. PMLR, 2021.
</div>
</section>
</section>
<section>
<section>
<h1>Direct search<br />for interpretable policies</h1>
<br />
<h3>
Get reward by looking for reward.
<br />
<br />
You don't have to use neural networks.
<br />
<br />
K.I.S.S.
</h3>
</section>
<section>
<h2>Functional trees as policies</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/aranha2.png" width="40%">
</div>
<div class="flex-child">
<img src="img/aranha1.png" width="50%">
</div>
</div>
Tree-based genetic programming to <b>maximize cumulative reward</b> on Cart Pole.
<div class="source">
Miranda, Ícaro, et al. "A Comparison Study Between Deep Learning and Genetic Programming
Application in Cart Pole Balancing Problem." 2018 IEEE Congress on Evolutionary Computation
(CEC). IEEE, 2018.
</div>
</section>
<section>
<h2>Interpretable Policies competitive with Deep RL</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/moo1.png" width="90%">
</div>
<div class="flex-child">
<img src="img/moo2.png" width="80%">
</div>
</div>
Simple policies with cumulative reward similar to Deep RL on control tasks in gym and pybullet.
<div class="source">
Videau, Mathurin, et al. "Multi-objective genetic programming for explainable reinforcement
learning." European Conference on Genetic Programming. 2022.
</div>
</section>
<section>
<h2>GP for eXplainable RL</h2>
<div class="flex-container">
<div class="flex-child">
<video height="400px" autoplay loop controls>
<source src="video/swingup.mp4" type="video/mp4">
</video>
</div>
<div class="flex-child">
<video height="400px" autoplay loop controls>
<source src="video/hopper.mp4" type="video/mp4">
</video>
</div>
</div>
<pre><code data-trim data-noescape>
def swingup(s):
return [s[4] + s[3]*6.614633680991087 - s[2] + np.exp(if_then_else(s[3]>-0.7571072906634332, s[1], \
15.013603569678889*s[1]))]
def hopper(s):
return [np.sin(np.exp(s[8])), -6.257060739725605*(s[7] + np.sin(s[3]+s[7])), \
np.sin(np.sin(s[7])-np.sin(s[8])-s[10]*(s[1]-np.log(abs(s[8]*s[3])+0.0001) - 5.860219777510614))]
</code></pre>
<div class="source">
Videau, Mathurin, et al. "Multi-objective genetic programming for explainable reinforcement
learning." European Conference on Genetic Programming. 2022.
</div>
</section>
<section>
<h2>Linear GP</h2>
<div class="flex-container">
<div class="flex-child">
<img src="img/lgp1.png" width="30%">
</div>
<div class="flex-child">
<img src="img/lgp2.png" width="80%">
</div>
</div>
<b>Register-based</b> genetic programming resulting in functional <b>graphs</b>.
<br />
Very simple to implement and uses standard GA.
<div class="source">
Kantschik, Wolfgang, and Wolfgang Banzhaf. "Linear-graph GP-a new GP structure." European
Conference on Genetic Programming. Berlin, Heidelberg: Springer Berlin Heidelberg, 2002.
<br />
Brameier, Markus, Wolfgang Banzhaf, and Wolfgang Banzhaf. Linear genetic programming. Vol. 1.
New York: Springer, 2007.
</div>
</section>
<section>
<h2>Programs as Policies</h2>
<img src="img/leaps.png" width="70%">
<br />
Program synthesis using continuous optimization over a latent program space.
<div class="source">
Trivedi, Dweep, et al. "Learning to synthesize programs as interpretable and generalizable
policies." Advances in neural information processing systems 34 (2021): 25146-25163.
</div>
</section>
<section>
<h2>Weight Agnostic Neural Networks</h2>
<div class="flex-container">
<div class="flex-child">
<video height="400px" autoplay loop controls>
<source src="https://weightagnostic.github.io/assets/mp4/square_biped.mp4"
type="video/mp4">
</video>
<img src="img/wann_left.png" width="40%">
</div>
<div class="flex-child">
<video height="400px" autoplay loop controls>
<source src="https://weightagnostic.github.io/assets/mp4/square_racer.mp4"
type="video/mp4">
</video>
<img src="img/wann_right.png" width="40%">
</div>
<br />
</div>
NEAT network optimization with single weight parameter shared across full network.<br />
<b>Parameter-free</b> functional graph.
<div class="source">
Gaier, Adam, and David Ha. "Weight agnostic neural networks." Advances in neural information
processing systems 32 (2019).
</div>
</section>
</section>
<section>
<section>
<h2>Cartesian Genetic Programming</h2>
<img class="plain" src="img/cgp/cgp.png">
<br />
A <b>graph-based</b> GP method based on functional node indexing with Cartesian coordinates.
<div class="source">
Miller, Julian Francis. "Cartesian genetic programming: its status and future." Genetic
Programming and Evolvable Machines 21 (2020): 129-168.
</div>
</section>
<section>
<h2>Floating Point Cartesian Genetic Programming</h2>
<img class="plain" src="img/cgp/fp_cgp.png" width="50%">
<br />
Modern CGP uses a <b>linear representation</b> and a single index per node.
<div class="source">
Wilson, Dennis G., et al. "Positional cartesian genetic programming." arXiv preprint
arXiv:1810.04119 (2018).
</div>
</section>
<section data-background-iframe="halfcheetah.html" data-background-interactive>
<h2>Graph-based policies</h2>
<br />
<br />
<br />
<br />
<br />
<img src="img/halfcheetah_cgp.png" width="20%">
<br />
Reward of 7000, similar to PPO and SAC.
</section>
<section data-background-iframe="hopper.html" data-background-interactive>
<h2>Graph-based policies</h2>
<br />
<br />
<br />
<br />
<br />
<br />
<img src="img/hopper_cgp.png" width="10%">
<br />
Reward of 20000 - 40000 (but no termination criteria)
</section>
<section>
<h2>Critical applications: Airplane taxi</h2>
<img src="img/l2f_taxi_path.png" width="40%">
<img src="img/rl-taxi2.png" width="40%">
<div class="source">
Wilson D, et al. "Learning Robust and Readable Control Laws for Aircraft Taxi Control." Under
review.
</div>
</section>
<section>
<h2>Critical applications: Airplane taxi</h2>
<img src="img/taxi-result.png" width="10%">
<img src="img/benchmark-convergence.png" width="30%">
<img src="img/attol-centerlne distance.png" width="30%">
<br />
Policy:
<br />
<img src="img/equationCGP.png" width="30%">
<div class="source">
Wilson D, et al. "Learning Robust and Readable Control Laws for Aircraft Taxi Control." Under
review.
</div>
</section>
</section>
<section>
<section data-background-video="video/asteroids.mp4" data-background-color="#000000">
<h2>Arcade Learning Environment</h2>
Bellemare, Marc G., et al. "The arcade learning environment: An evaluation platform for general
agents."<br/> Journal of Artificial Intelligence Research 47 (2013): 253-279.
<br />
<br />
Standard benchmark for visual control
<br />
<br />
Used in:
<br/>
<ul>
<li>HyperNEAT [Hausknecht et al., 2012]</li>
<li>DQN [Mnih et al., 2015]</li>
<li>A3C [Mnih et al., 2016]</li>
<li>Tangled Problem Graphs [Kelly and Heywood, 2017]</li>
<li>CGP [Wilson et al., 2018]</li>
<li>Go-explore [Ecoffet et al., 2019]</li>
<li>Flare [Shang et al., 2021]</li>
</ul>
</section>
<section>
<h2>Control with convolutional neural network policy</h2>
<img class="plain" src="img/dqn.png">
<br />
Policy network maps pixel space representation to small action distribution
<div class="source">
Mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning." Nature
518.7540 (2015): 529-533.
</div>
</section>
<section>
<h2>Explaining Atari agents</h2>
<img class="plain" src="img/pong_saliency.png">
<img class="plain" src="img/spaceinvaders_saliency.png">
<br />
Attention maps can show <b>what</b> information a network is using.
<br />
Observation and expertise can, in some cases, demonstrate <b>how</b> that information informs the
decision.
<div class="source">
Greydanus, Samuel, et al. "Visualizing and understanding atari agents." International conference
on machine learning. PMLR, 2018.
</div>
</section>
<section>
<h2>CGP for Atari playing</h2>
<img class="plain" src="img/scheme.png">
<div class="source">
Wilson, Dennis G., et al. "Evolving simple programs for playing Atari games." Proceedings of the
genetic and evolutionary computation conference. 2018.
</div>
</section>
<section>
<h6>Centipede</h6>
<video height="400px" autoplay loop controls>
<source src="video/centipede.mp4" type="video/mp4">
</video>
<img class="plain" src="img/centipede_graph.png" height="400px">
<br />
<br />
<small>
<table>
<thead>
<tr>
<th>Human</th>
<th>Double</th>
<th>DQN</th>
<th>Prioritized</th>
<th>A3C:FF</th>
<th>A3C:LSTM</th>
<th>TPG</th>
<th>HyperNEAT</th>
<th>CGP</th>
</tr>
</thead>
<tbody>
<tr>
<td>11963</td>
<td>3853.5</td>
<td>4881</td>
<td>3421.9</td>
<td>3755.8</td>
<td>1997</td>
<td>34731.7</td>
<td>25275.2</td>
<td>24708</td>
</tr>
</tbody>
</table>
</small>
<div class="source">
Wilson, Dennis G., et al. "Evolving simple programs for playing Atari games." Proceedings of the
genetic and evolutionary computation conference. 2018.
</div>
</section>
<section>
<h2>Kung Fu Master</h2>
<video height="400px" autoplay loop controls>
<source src="video/kung_fu_master.mp4" type="video/mp4">
</video>
<img class="plain" src="img/kung_fu_master_graph.png" height="400px">
<br />
<br />
<small>
<table>
<thead>
<tr>
<th>Human</th>
<th>Double</th>
<th>DQN</th>
<th>Prioritized</th>
<th>A3C:FF</th>
<th>A3C:LSTM</th>
<th>TPG</th>
<th>HyperNEAT</th>
<th>CGP</th>
</tr>
</thead>
<tbody>
<tr>
<td>22736</td>
<td>30207</td>
<td>24288</td>
<td>31244</td>
<td>28819</td>
<td>40835</td>
<td></td>
<td>7720</td>
<td>57400</td>
</tr>
</tbody>
</table>
</small>
<div class="source">
Wilson, Dennis G., et al. "Evolving simple programs for playing Atari games." Proceedings of the
genetic and evolutionary computation conference. 2018.
</div>
</section>
<section>
<h2>Boxing</h2>
<video height="400px" autoplay loop controls>
<source src="video/boxing.mp4" type="video/mp4">
</video>
<img class="plain" src="img/boxing_graph.png" height="400px">
<br />
<br />
<small>
<table>
<thead>
<tr>
<th>Human</th>
<th>Double</th>
<th>DQN</th>
<th>Prioritized</th>
<th>A3C:FF</th>
<th>A3C:LSTM</th>
<th>TPG</th>
<th>HyperNEAT</th>
<th>CGP</th>
</tr>
</thead>
<tbody>
<tr>
<td>4.3</td>
<td>73.5</td>
<td>77.3</td>
<td>68.6</td>
<td>59.8</td>
<td>37.3</td>
<td></td>
<td>16.4</td>
<td>38.4</td>
</tr>
</tbody>
</table>
</small>
<div class="source">
Wilson, Dennis G., et al. "Evolving simple programs for playing Atari games." Proceedings of the
genetic and evolutionary computation conference. 2018.
</div>
</section>
<section>
<h2>Interpretability in high-dimensional data</h2>
<img src="img/self_driving_1.png" width="50%">
<ul>
<li>Break the data down into lower dimension</li>
<li>Use explanability methods like attention maps</li>
<li>Construct an interpretable analysis pipeline using adapted functions</li>
<li>Split the problem into different interpretable parts</li>
</ul>
<div class="source">
Atakishiyev, Shahin, et al. "Explainable artificial intelligence for autonomous driving: A
comprehensive overview and field guide for future research directions." arXiv preprint
arXiv:2112.11561 (2021).
</div>
</section>
<section>
<h2>TPG: Decimal Feature Grid</h2>
<img class="plain" src="img/tpg.png">
<br />
Break the data down into lower dimension: small grid with discretized color palette
<div class="source">
Kelly, Stephen, and Malcolm I. Heywood. "Emergent tangled graph representations for Atari game
playing agents."<br /> Genetic Programming: 20th European Conference, EuroGP 2017, Amsterdam,
The Netherlands, April 19-21, 2017, Proceedings 20. Springer International Publishing, 2017.
</div>
</section>
<section>
<h2>TPG: Multi-Task Learning</h2>
<img class="plain" src="img/tpg_multi.png" height="500">
<br />
Interpretable graphs inherently allow for understanding <b>mutual information</b> between control
policies.
<div class="source">
Kelly, Stephen, and Malcolm I. Heywood. "Multi-task learning in atari video games with emergent
tangled program graphs." Proceedings of the Genetic and Evolutionary Computation Conference.
2017.
</div>
</section>
<section>
<h2>Interpretable CGP Encoder Controller</h2>
<img class="plain" src="img/icec1.png" height="500">
<br />
Construct an interpretable analysis pipeline using adapted functions.
<br />
Split the problem into different interpretable parts.
<div class="source">
Lecarpentier, Erwan, et al. "Cartesian Genetic Programming for Learning Interpretable Agents
Playing Atari Games." Under review.
</div>
</section>
<section>
<h2>Interpretable CGP Encoder Controller</h2>
<img class="plain" src="img/icec2.png" width="30%">
<br />
Bowling:
<br />
hit action button when character is in 5,2
<br />
hit down direction when ball is launched
<br />
hit up direction when ball is in 4,3 or 4,5.
<div class="source">
Lecarpentier, Erwan, et al. "Cartesian Genetic Programming for Learning Interpretable Agents
Playing Atari Games." Under review.
</div>
</section>
</section>
<section>
<section>
<h2>Takeaways</h2>
<ul>
<li>Explainable policies are necessary for critical applications</li>
<li>Interpretable policy representations give naturally to explanations</li>
<li>Evolutionary policy search is well-suited to interpretable representations</li>
<li>Many representations exist: trees, graphs, programs</li>
<li>Not a very active domain and requires more comparison between methods</li>
<li>High-dimensional data like visual representations still difficult</li>
</ul>
<br/>
<video height="400px" autoplay loop controls>
<source src="video/qbert.mp4" type="video/mp4">
</video>
<br/>
Thank you!
</section>
</section>
<div id="footer-container" style="display:none;">
<div id="footer">
Evolutionary Methods for Interpretable Control, GECCO 2023
<br />
<a
href="https://github.com/d9w/interpretable_control/">https://github.com/d9w/interpretable_control</a>
<br />
Dennis G. Wilson
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/"><img
alt="Creative Commons License" style="border-width:0"
src="https://i.creativecommons.org/l/by-sa/4.0/80x15.png" /></a>
</div>
</div>
</div>
<script src="reveal/js/reveal.js"></script>
<script src="reveal/js/jquery-3.5.0.min.js"></script>
<script>
Reveal.initialize({
width: "90%",
height: "90%",
slideNumber: 'c',
transition: 'none',
progress: true,
hash: true,
center: false,
dependencies: [
{ src: 'reveal/plugin/markdown/marked.js' },
{ src: 'reveal/plugin/markdown/markdown.js' },
{ src: 'reveal/plugin/math/math.js' },
{ src: 'reveal/plugin/highlight/highlight.js' },
{ src: 'reveal/plugin/notes/notes.js', async: true }
]
});
var footer = $('#footer-container').html();
$('div.reveal').append(footer);
</script>
</body>
</html>