-
Notifications
You must be signed in to change notification settings - Fork 6
/
neuroevolution.tex
398 lines (342 loc) · 12.2 KB
/
neuroevolution.tex
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
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% MUW Presentation
% LaTeX Template
% Version 1.0 (27/12/2016)
%
% License:
% CC BY-NC-SA 4.0 (http://creativecommons.org/licenses/by-nc-sa/3.0/)
%
% Created by:
% Nicolas Ballarini, CeMSIIS, Medical University of Vienna
% http://statistics.msi.meduniwien.ac.at/
%
% Customized for UAH by:
% David F. Barrero, Departamento de Automática, UAH
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\documentclass[10pt,compress]{beamer} % Change 10pt to make fonts of a different size
\mode<presentation>
\usepackage[spanish]{babel}
\usepackage{fontspec}
\usepackage{tikz}
\usepackage{etoolbox}
\usepackage{xcolor}
\usepackage{xstring}
\usepackage{listings}
% Custom packages
\usepackage{tikz}
\usepackage{standalone}
\def\layersep{2.5cm}
\usetikzlibrary{matrix,chains,positioning,decorations.pathreplacing,arrows,tikzmark}
\usetheme{UAH}
\usecolortheme{UAH}
\setbeamertemplate{navigation symbols}{}
\setbeamertemplate{caption}[numbered]
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Presentation Info
\title[Neuroevolution]{Neuroevolution}
\author{\asignatura\\\carrera}
\institute{}
\date{Departamento de Automática}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Descomentar para habilitar barra de navegación superior
\setNavigation
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Configuración de logotipos en portada
%% Opacidad de los logotipos
\newcommand{\opacidad}{1}
%% Descomentar para habilitar logotipo en pié de página de portada
\renewcommand{\logoUno}{Images/isg.png}
%% Descomentar para habilitar logotipo en pié de página de portada
%\renewcommand{\logoDos}{Images/CCLogo.png}
%% Descomentar para habilitar logotipo en pié de página de portada
%\renewcommand{\logoTres}{Images/ALogo.png}
%% Descomentar para habilitar logotipo en pié de página de portada
%\renewcommand{\logoCuatro}{Images/ELogo.png}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% FOOTLINE
%% Comment/Uncomment the following blocks to modify the footline
%% content in the body slides.
%% Option A: Title and institute
\footlineA
%% Option B: Author and institute
%\footlineB
%% Option C: Title, Author and institute
%\footlineC
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Use this block for a blue title slide with modified footline
{\titlepageBlue
\begin{frame}
\titlepage
\end{frame}
}
\begin{frame}[plain]{}
\begin{block}{Objective}
\begin{itemize}
\item Fusion ANN and Evolutionary Algorithms
\item Identify application areas of Neuroevolution in Robotics
\end{itemize}
\end{block}
\begin{block}{Bibliography}
\begin{enumerate}
\item A. Tettamanzi, M. Tomassini. \textit{Soft Computing. Integrating Evolutionary, Neural, and Fuzzy Systems}. Springer-Verlag. 2001
\item D. Floreano, P. D\"urr, C. Mattiussi. \textit{Neuroevolution: from architectures to learning}. Evolutionary Intelligence, Vol. 1, No. 1, pags. 47-62. Springer-Verlag. 2008.
\item S. Risi, J. Togelius. \textit{Neuroevolution in Games: State of the Art and Open Challenges}. IEEE Trans. on Computational Intelligence and AI in Games. Vol 9, No. 1. 2017
\end{enumerate}
\end{block}
\end{frame}
{
\disableNavigation{white}
\begin{frame}[shrink]{Table of Contents}
\frametitle{Table of Contents}
\tableofcontents
% You might wish to add the option [pausesections]
\end{frame}
}
\section{Introduction}
\subsection{Motivation}
\begin{frame}{Introduction}{Motivation}
Problems with traditional learning algorithms
\begin{itemize}
\item Fixed topology
\item Local mimina and other learning limitations
\end{itemize}
In Nature ...
\begin{itemize}
\item Global neural system architecture given by evolution
\item Details (synapsys) given by learning
\end{itemize}
EAs good avoiding local maxima and searching complex search spaces
\end{frame}
\subsection{Definition}
\begin{frame}{Introduction}{Definition (I)}
\begin{columns}
\column{.80\textwidth}
\vspace{-0.3cm}
\begin{block}{Neuroevolution (NE)}
\textit{NE refers to the generation of artificial neural networks (their connections weights and/or topology) using evolutionary algorithms}
\vspace{-0.2cm}
\begin{flushright} Risi and Togelius\end{flushright}
\end{block}
\end{columns}
NE features
\begin{itemize}
\item Record-beating performance
\item Broad applicability
\item Scalability
\item Diversity
\item Open-ended learning
\end{itemize}
Problem: NE does not provide explicative models
\end{frame}
\begin{frame}{Introduction}{Definition (II)}
When ANN meets EAs
\begin{itemize}
\item Evolve weights
\item Evolve topology
\end{itemize}
Key elements to take into account
\begin{itemize}
\item Evaluation (straight or hybrid) (not accepted terms)
\item Representation (direct or indirect)
\end{itemize}
\end{frame}
\section{ANN evaluation}
\subsection{Straight approach}
\begin{frame}{ANN evaluation}{Straight approach}
Straight: Build ANN and assess it
\begin{itemize}
\item Cumulative error on test set
\item Simulate robot behavior
\item Observe robot behavior
\end{itemize}
Some problems
\begin{itemize}
\item Pretty slow
\item Does not explot gradient if available
\end{itemize}
\end{frame}
\subsection{Hybrid approach}
\begin{frame}{ANN evaluation}{Hybrid approach}
Hybrid: Join evolution and learning
\begin{itemize}
\item EAs have good exploration
\item Training has good explotation
\item Gradient-based methods sensible to initial weights
\end{itemize}
Evolve a population of ANN, then train them (backprop, ...)
\begin{itemize}
\item Trainning does not change genotype
\end{itemize}
\begin{center}
\includegraphics[width=0.4\linewidth]{figs/hybrid.png}\\
\tiny{A. Tettamanzi, M. Tomassini. \textit{Soft Computing. Integrating Evolutionary, Neural, and Fuzzy Systems}. Springer-Verlag. 2001}
\end{center}
\end{frame}
\section{Direct codification}
\subsection{Direct codification of ANNs}
\begin{frame}{Direct codification}{Direct codification of ANNs}
\begin{columns}
\column{.50\textwidth}
Each gene represents a weight
\begin{itemize}
\item Fixed topology
\end{itemize}
Can be used any EA
\begin{itemize}
\item Typically, GA or ES
\end{itemize}
Several complex neuroevolution specific codifications
\column{.50\textwidth}
\input{figs/direct.tex}
\end{columns}
\end{frame}
\subsection{Alternative direct codification of ANNs}
\begin{frame}{Direct codification}{Alternative direct codification of ANNs}
\tikzstyle{every picture}+=[remember picture]
\begin{columns}
\column{.50\textwidth}
\input{figs/alternative.tex}
\column{.50\textwidth}
\begin{table}[]
\centering
\begin{tabular}{l|llllll}
& 1 & 2 & 3 & 4 & 5 & 6 \\\hline
1 & 0 & 0 & 0 & \tikzmark{four}1 & \tikzmark{five}1 & 0 \\
2 & 0 & 0 & 0 & 1 & 0 & 1 \\
3 & 0 & 0 & 0 & 0 & 1 & 0 \\
4 & 0 & 0 & 0 & 0 & 0 & 1 \\
5 & 0 & 0 & 0 & 0 & 0 & 1 \\
6 & 0 & 0 & 0 & 0 & 0 & 0
\end{tabular}
\end{table}
\begin{tikzpicture}[overlay, remember picture]
\draw <2-> [->, blue!50, in=90, out=120] ({pic cs:four}) to (aux14);
\draw <2-> [->, blue!50, in=90, out=120] ({pic cs:five}) to (aux15);
\end{tikzpicture}
\end{columns}
\end{frame}
\subsection{Permutation problem}
\begin{frame}{Direct codification}{Permutation problem}
Permutation problem (also known as compeling convenions)
\begin{itemize}
\item Multiple genotypes with same phenotype
\end{itemize}
\input{figs/direct.tex}
\input{figs/direct-inverse.tex}
\end{frame}
\subsection{Symbolic, adaptive, neuro-evolution (SANE)}
\begin{frame}{Direct codification}{Symbolic, adaptive, neuro-evolution (SANE)}
\begin{columns}
\column{.50\textwidth}
SANE evolves single neurons
\begin{itemize}
\item Population of neurons
\item Connections and weights
\item Fixed topology: One hidden layer
\end{itemize}
\column{.50\textwidth}
Evaluation
\begin{enumerate}
\item Build random ANNs with sampled neurons
\item Compute ANNs fitness
\item Fitness of a neuron is the average fitness of all the ANNs it has participated in
\end{enumerate}
\end{columns}
\begin{center}
\includegraphics[width=0.5\linewidth]{figs/sane.png}\\
\tiny{ D. Floreano, P. D\"urr, C. Mattiussi. \textit{Neuroevolution: from architectures to learning}. Evolutionary Intelligence, Vol. 1, No. 1, pags. 47-62. Springer-Verlag. 2008.}
\end{center}
\end{frame}
\subsection{Neuro-evolution of Augmenting Topologies (NEAT)}
\begin{frame}{Direct codification}{Neuro-evolution of Augmenting Topologies (NEAT)}
\begin{columns}
\column{.50\textwidth}
Quite used NE algorithm
\begin{itemize}
\item Weights and \textit{topologies}
\item Grows ANN complexity
\end{itemize}
Two types of genes
\begin{itemize}
\item Nodes and connections
\end{itemize}
Genetic operators
\begin{itemize}
\item Meaningful crossover
\item Gene enable/disable mutation
\item Gene insert operator
\end{itemize}
\href{https://www.youtube.com/watch?v=3JiNC6vw8zE}{(Video Torcs)}
\href{https://www.youtube.com/watch?v=tmltm0ZHkHw}{(Video Mario)}
\column{.50\textwidth}
\begin{center}
\includegraphics[width=\linewidth]{figs/neat.png}\\
\tiny{ D. Floreano, P. D\"urr, C. Mattiussi. \textit{Neuroevolution: from architectures to learning}. Evolutionary Intelligence, Vol. 1, No. 1, pags. 47-62. Springer-Verlag. 2008.}
\end{center}
\end{columns}
\end{frame}
%\subsection{Covariance Matrix Adaptation (CMA-ES)}
\section{Indirect codification}
\subsection{Introduction}
\begin{frame}{Indirect codification}{Introduction}
Problems with direct codification
\begin{itemize}
\item Scalability
\item No reuse
\end{itemize}
Indirect encoding try to grow networks
\begin{itemize}
\item Evolve generation rules instead of individual weights
\item Try to reuse basic building blocks
\item Closer to biological systems
\end{itemize}
\end{frame}
\subsection{Kitano's method}
\begin{frame}[fragile]{Indirect codification}{Kitano's method (I)}
\begin{columns}
\column{.40\textwidth}
Kitano used rewriting rules
\begin{itemize}
\item Terminals, a symbol
\item Non-terminals, a rewriting rule
\end{itemize}
\column{.60\textwidth}
\begin{exampleblock}{Grammar example}
\small{
\texttt{
<digit> $\longrightarrow$ 0|1|2|3|4|5|6|7|8|9\\
<number> $\longrightarrow$ <digit>\\
<number> $\longrightarrow$ <number><digit>}
}
\end{exampleblock}
\end{columns}
\bigskip
Rather standard GA evolve rules
\begin{itemize}
\item Fitness proportionale, elitism, variable mutation rate, single crossover
\item Evaluation of the network trained with backpropagation
\end{itemize}
Cromosomes composed by
\begin{itemize}
\item Fixed and evolvable rules
\end{itemize}
\end{frame}
\begin{frame}[plain]{Indirect codification}{Kitano's method (II)}
\begin{center}
\includegraphics[width=\linewidth]{figs/kitano.png}\\
\tiny{D. Floreano, P. D\"urr, C. Mattiussi. \textit{Neuroevolution: from architectures to learning}. Evolutionary Intelligence, Vol. 1, No. 1, pags. 47-62. Springer-Verlag. 2008.}
\end{center}
\end{frame}
\begin{frame}[plain]{Indirect codification}{Kitano's method (III)}
\begin{center}
\includegraphics[width=0.7\linewidth]{figs/kitanoCodification.png}\\
\tiny{H. Kitano. \textit{Designing Neural Networks Using Genetic Algorithms with Graph Generation System}. Complex Systems 4: 461-476. 1990.}
\end{center}
\end{frame}
\end{document}