Strict Standards: Declaration of action_plugin_archiveupload::register() should be compatible with DokuWiki_Action_Plugin::register($controller) in /www/htdocs/v108444/lib/plugins/archiveupload/action.php on line 19
Strict Standards: Declaration of cache_instructions::retrieveCache() should be compatible with cache::retrieveCache($clean = true) in /www/htdocs/v108444/inc/cache.php on line 289
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/inc/auth.php on line 295
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 109
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 114
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 115
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 116
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/inc/pageutils.php on line 545
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/inc/pageutils.php on line 546
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 136
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 143
Warning: Cannot modify header information - headers already sent by (output started at /www/htdocs/v108444/lib/plugins/archiveupload/action.php:19) in /www/htdocs/v108444/lib/exe/fetch.php on line 175
%PDF-1.4
%
1 0 obj
<< /S /GoTo /D (section*.3) >>
endobj
4 0 obj
(Abbildungsverzeichnis)
endobj
5 0 obj
<< /S /GoTo /D (section*.5) >>
endobj
8 0 obj
(Tabellenverzeichnis)
endobj
9 0 obj
<< /S /GoTo /D (chapter.1) >>
endobj
12 0 obj
(1 Einleitung)
endobj
13 0 obj
<< /S /GoTo /D (section.1.1) >>
endobj
16 0 obj
(1.1 Handlungslernen auf Basis visueller Wahrnehmungen)
endobj
17 0 obj
<< /S /GoTo /D (section.1.2) >>
endobj
20 0 obj
(1.2 Biologische Systeme)
endobj
21 0 obj
<< /S /GoTo /D (section.1.3) >>
endobj
24 0 obj
(1.3 Klassische Herangehensweise)
endobj
25 0 obj
<< /S /GoTo /D (section.1.4) >>
endobj
28 0 obj
(1.4 Integrierte Herangehensweise)
endobj
29 0 obj
<< /S /GoTo /D (section.1.5) >>
endobj
32 0 obj
(1.5 Zielsetzung und Validierung)
endobj
33 0 obj
<< /S /GoTo /D (section.1.6) >>
endobj
36 0 obj
(1.6 Zweistufiger, integrierter Ansatz: ``Deep Fitted Q'')
endobj
37 0 obj
<< /S /GoTo /D (section.1.7) >>
endobj
40 0 obj
(1.7 Stufe 1: Wahrnehmung)
endobj
41 0 obj
<< /S /GoTo /D (section.1.8) >>
endobj
44 0 obj
(1.8 Stufe 2: Strategielernen)
endobj
45 0 obj
<< /S /GoTo /D (section.1.9) >>
endobj
48 0 obj
(1.9 Vorgehen und Aufbau der Arbeit)
endobj
49 0 obj
<< /S /GoTo /D (chapter.2) >>
endobj
52 0 obj
(2 Grundlagen)
endobj
53 0 obj
<< /S /GoTo /D (section.2.1) >>
endobj
56 0 obj
(2.1 Tiefes Lernen)
endobj
57 0 obj
<< /S /GoTo /D (subsection.2.1.1) >>
endobj
60 0 obj
(2.1.1 K\374nstliche neuronale Netze)
endobj
61 0 obj
<< /S /GoTo /D (subsection.2.1.2) >>
endobj
64 0 obj
(2.1.2 Mehrschichtige Perzeptronnetze)
endobj
65 0 obj
<< /S /GoTo /D (subsection.2.1.3) >>
endobj
68 0 obj
(2.1.3 Autoencoder)
endobj
69 0 obj
<< /S /GoTo /D (subsection.2.1.4) >>
endobj
72 0 obj
(2.1.4 Training tiefer Autoencoder)
endobj
73 0 obj
<< /S /GoTo /D (section.2.2) >>
endobj
76 0 obj
(2.2 Reinforcement Lernen)
endobj
77 0 obj
<< /S /GoTo /D (subsection.2.2.1) >>
endobj
80 0 obj
(2.2.1 Markovsche Entscheidungsprozesse)
endobj
81 0 obj
<< /S /GoTo /D (subsection.2.2.2) >>
endobj
84 0 obj
(2.2.2 Diskontierte und nicht diskontierte Problemstellungen)
endobj
85 0 obj
<< /S /GoTo /D (subsection.2.2.3) >>
endobj
88 0 obj
(2.2.3 Optimale Wertfunktion und optimale Strategie)
endobj
89 0 obj
<< /S /GoTo /D (subsection.2.2.4) >>
endobj
92 0 obj
(2.2.4 Wertiteration)
endobj
93 0 obj
<< /S /GoTo /D (subsection.2.2.5) >>
endobj
96 0 obj
(2.2.5 Q-Lernen)
endobj
97 0 obj
<< /S /GoTo /D (subsection.2.2.6) >>
endobj
100 0 obj
(2.2.6 Approximatives Reinforcement Lernen)
endobj
101 0 obj
<< /S /GoTo /D (subsection.2.2.7) >>
endobj
104 0 obj
(2.2.7 Batch Reinforcement Lernen)
endobj
105 0 obj
<< /S /GoTo /D (subsection.2.2.8) >>
endobj
108 0 obj
(2.2.8 Fitted Q-Iteration)
endobj
109 0 obj
<< /S /GoTo /D (subsection.2.2.9) >>
endobj
112 0 obj
(2.2.9 Batch Reinforcement Lernen mit Exploration)
endobj
113 0 obj
<< /S /GoTo /D (chapter.3) >>
endobj
116 0 obj
(3 DFQ: Optimierendes Lernen auf hochdimensionalen Observationen)
endobj
117 0 obj
<< /S /GoTo /D (section.3.1) >>
endobj
120 0 obj
(3.1 Motivation)
endobj
121 0 obj
<< /S /GoTo /D (section.3.2) >>
endobj
124 0 obj
(3.2 Struktur der Interaktion mit der Umgebung)
endobj
125 0 obj
<< /S /GoTo /D (section.3.3) >>
endobj
128 0 obj
(3.3 Training des Agenten)
endobj
129 0 obj
<< /S /GoTo /D (subsection.3.3.1) >>
endobj
132 0 obj
(3.3.1 Rahmenablauf ohne Exploration)
endobj
133 0 obj
<< /S /GoTo /D (subsection.3.3.2) >>
endobj
136 0 obj
(3.3.2 Rahmenablauf mit Exploration)
endobj
137 0 obj
<< /S /GoTo /D (subsection.3.3.3) >>
endobj
140 0 obj
(3.3.3 Schnittstellen und R\374ckkopplung zwischen den beiden Stufen)
endobj
141 0 obj
<< /S /GoTo /D (section.3.4) >>
endobj
144 0 obj
(3.4 Deep Fitted Q-Iteration \(DFQ\))
endobj
145 0 obj
<< /S /GoTo /D (section.3.5) >>
endobj
148 0 obj
(3.5 Effizienzsteigerung beim Generationswechsel)
endobj
149 0 obj
<< /S /GoTo /D (subsection.3.5.1) >>
endobj
152 0 obj
(3.5.1 Sofortige Neuberechnung einer Wertfunktion)
endobj
153 0 obj
<< /S /GoTo /D (subsection.3.5.2) >>
endobj
156 0 obj
(3.5.2 ``\334bersetzen'' der Wertfunktion)
endobj
157 0 obj
<< /S /GoTo /D (subsection.3.5.3) >>
endobj
160 0 obj
(3.5.3 Speichern der Transitionen im Merkmalsraum)
endobj
161 0 obj
<< /S /GoTo /D (subsection.3.5.4) >>
endobj
164 0 obj
(3.5.4 Vermeidung von Generationswechseln)
endobj
165 0 obj
<< /S /GoTo /D (section.3.6) >>
endobj
168 0 obj
(3.6 Zusammenfassung)
endobj
169 0 obj
<< /S /GoTo /D (chapter.4) >>
endobj
172 0 obj
(4 Zur Konvergenz und Konsistenz des DFQ-Algorithmus)
endobj
173 0 obj
<< /S /GoTo /D (section.4.1) >>
endobj
176 0 obj
(4.1 Grundlagen)
endobj
177 0 obj
<< /S /GoTo /D (subsection.4.1.1) >>
endobj
180 0 obj
(4.1.1 Konvergenz der approximativen Wertiteration)
endobj
181 0 obj
<< /S /GoTo /D (subsection.4.1.2) >>
endobj
184 0 obj
(4.1.2 Qualit\344t der L\366sung der approximativen Wertiteration)
endobj
185 0 obj
<< /S /GoTo /D (subsection.4.1.3) >>
endobj
188 0 obj
(4.1.3 Konvergenz des Lernens auf Basis einer Stichprobe)
endobj
189 0 obj
<< /S /GoTo /D (subsection.4.1.4) >>
endobj
192 0 obj
(4.1.4 Stochastische Konsistenz des Lernens auf Stichproben)
endobj
193 0 obj
<< /S /GoTo /D (subsection.4.1.5) >>
endobj
196 0 obj
(4.1.5 Zusammenfassung)
endobj
197 0 obj
<< /S /GoTo /D (section.4.2) >>
endobj
200 0 obj
(4.2 Fitted Q-Iterations ist Variante des KADP-Verfahrens)
endobj
201 0 obj
<< /S /GoTo /D (section.4.3) >>
endobj
204 0 obj
(4.3 Situation in DFQ)
endobj
205 0 obj
<< /S /GoTo /D (section.4.4) >>
endobj
208 0 obj
(4.4 Zur Konvergenz in der inneren Schleife)
endobj
209 0 obj
<< /S /GoTo /D (subsection.4.4.1) >>
endobj
212 0 obj
(4.4.1 Definition korrespondierender Wertfunktionen)
endobj
213 0 obj
<< /S /GoTo /D (subsection.4.4.2) >>
endobj
216 0 obj
(4.4.2 Konstruktion eines korrespondierenden Zufallsoperators)
endobj
217 0 obj
<< /S /GoTo /D (subsection.4.4.3) >>
endobj
220 0 obj
(4.4.3 Herleitung der Konvergenz zu einem eindeutigen Fixpunkt)
endobj
221 0 obj
<< /S /GoTo /D (subsection.4.4.4) >>
endobj
224 0 obj
(4.4.4 Ergebnisse f\374r nicht diskontierte MDPs und die Fehlerabsch\344tzung)
endobj
225 0 obj
<< /S /GoTo /D (section.4.5) >>
endobj
228 0 obj
(4.5 Zur Konsistenz der \344u\337eren Schleife)
endobj
229 0 obj
<< /S /GoTo /D (subsection.4.5.1) >>
endobj
232 0 obj
(4.5.1 Stochastische Konsistenz \374ber Explorationsphasen hinweg)
endobj
233 0 obj
<< /S /GoTo /D (subsection.4.5.2) >>
endobj
236 0 obj
(4.5.2 Einfluss der Generationswechsel)
endobj
237 0 obj
<< /S /GoTo /D (section.4.6) >>
endobj
240 0 obj
(4.6 Diskussion der Ergebnisse)
endobj
241 0 obj
<< /S /GoTo /D (section.4.7) >>
endobj
244 0 obj
(4.7 Zusammenfassung)
endobj
245 0 obj
<< /S /GoTo /D (chapter.5) >>
endobj
248 0 obj
(5 Automatische Konstruktion von Merkmalsr\344umen in DFQ)
endobj
249 0 obj
<< /S /GoTo /D (section.5.1) >>
endobj
252 0 obj
(5.1 Motivation)
endobj
253 0 obj
<< /S /GoTo /D (section.5.2) >>
endobj
256 0 obj
(5.2 Training mit flachen Autoencodern)
endobj
257 0 obj
<< /S /GoTo /D (subsection.5.2.1) >>
endobj
260 0 obj
(5.2.1 In DFQ verwendeter Algorithmus)
endobj
261 0 obj
<< /S /GoTo /D (subsection.5.2.2) >>
endobj
264 0 obj
(5.2.2 Effiziente Implementierung)
endobj
265 0 obj
<< /S /GoTo /D (subsection.5.2.3) >>
endobj
268 0 obj
(5.2.3 Rezeptive Felder und Faltungskerne)
endobj
269 0 obj
<< /S /GoTo /D (subsection.5.2.4) >>
endobj
272 0 obj
(5.2.4 Parameter und Modellauswahl)
endobj
273 0 obj
<< /S /GoTo /D (section.5.3) >>
endobj
276 0 obj
(5.3 Empirische Evaluation)
endobj
277 0 obj
<< /S /GoTo /D (subsection.5.3.1) >>
endobj
280 0 obj
(5.3.1 Kontinuierliche Grid-World)
endobj
281 0 obj
<< /S /GoTo /D (subsection.5.3.2) >>
endobj
284 0 obj
(5.3.2 Grundlegende Ergebnisse)
endobj
285 0 obj
<< /S /GoTo /D (subsection.5.3.3) >>
endobj
288 0 obj
(5.3.3 Vergleich mit der Hauptkomponentenanalyse)
endobj
289 0 obj
<< /S /GoTo /D (subsection.5.3.4) >>
endobj
292 0 obj
(5.3.4 R\374ckw\344rtspropagieren des Gradienten in den Encoder)
endobj
293 0 obj
<< /S /GoTo /D (subsection.5.3.5) >>
endobj
296 0 obj
(5.3.5 Verkleinerung der Trainingsmenge f\374hrt zu Problemen)
endobj
297 0 obj
<< /S /GoTo /D (subsection.5.3.6) >>
endobj
300 0 obj
(5.3.6 Faltungskerne erleichtern das Training)
endobj
301 0 obj
<< /S /GoTo /D (subsection.5.3.7) >>
endobj
304 0 obj
(5.3.7 Anwendung auf reale Bilddaten)
endobj
305 0 obj
<< /S /GoTo /D (section.5.4) >>
endobj
308 0 obj
(5.4 Diskussion und Zusammenfassung)
endobj
309 0 obj
<< /S /GoTo /D (chapter.6) >>
endobj
312 0 obj
(6 Adaptive Partitionierung des Merkmalsraums in DFQ)
endobj
313 0 obj
<< /S /GoTo /D (section.6.1) >>
endobj
316 0 obj
(6.1 Motivation)
endobj
317 0 obj
<< /S /GoTo /D (section.6.2) >>
endobj
320 0 obj
(6.2 Optimierung der Struktur eines Gitterapproximators)
endobj
321 0 obj
<< /S /GoTo /D (section.6.3) >>
endobj
324 0 obj
(6.3 ClusterRL-Algorithmus)
endobj
325 0 obj
<< /S /GoTo /D (subsection.6.3.1) >>
endobj
328 0 obj
(6.3.1 Lloyds Algorithmus \(alias k-means\))
endobj
329 0 obj
<< /S /GoTo /D (subsection.6.3.2) >>
endobj
332 0 obj
(6.3.2 ``Training'' des irregul\344ren Gitterapproximators)
endobj
333 0 obj
<< /S /GoTo /D (subsection.6.3.3) >>
endobj
336 0 obj
(6.3.3 Abfrage des irregul\344ren Gitterapproximators)
endobj
337 0 obj
<< /S /GoTo /D (subsection.6.3.4) >>
endobj
340 0 obj
(6.3.4 Implementierung)
endobj
341 0 obj
<< /S /GoTo /D (subsection.6.3.5) >>
endobj
344 0 obj
(6.3.5 Konvergenzverhalten)
endobj
345 0 obj
<< /S /GoTo /D (section.6.4) >>
endobj
348 0 obj
(6.4 Empirische Evaluation)
endobj
349 0 obj
<< /S /GoTo /D (subsection.6.4.1) >>
endobj
352 0 obj
(6.4.1 Mountain-Car)
endobj
353 0 obj
<< /S /GoTo /D (subsection.6.4.2) >>
endobj
356 0 obj
(6.4.2 Approximation der optimalen Zustands-Wertfunktion V*\(s\))
endobj
357 0 obj
<< /S /GoTo /D (subsection.6.4.3) >>
endobj
360 0 obj
(6.4.3 Erprobung der Trainingsvarianten)
endobj
361 0 obj
<< /S /GoTo /D (subsection.6.4.4) >>
endobj
364 0 obj
(6.4.4 Lernen auf Mannigfaltigkeit)
endobj
365 0 obj
<< /S /GoTo /D (subsection.6.4.5) >>
endobj
368 0 obj
(6.4.5 Zusammenfassung der Evaluation)
endobj
369 0 obj
<< /S /GoTo /D (section.6.5) >>
endobj
372 0 obj
(6.5 Deep Fitted Q-Iteration mit Clustern: DFQ-C)
endobj
373 0 obj
<< /S /GoTo /D (section.6.6) >>
endobj
376 0 obj
(6.6 Zusammenfassung)
endobj
377 0 obj
<< /S /GoTo /D (chapter.7) >>
endobj
380 0 obj
(7 Empirische Evaluation)
endobj
381 0 obj
<< /S /GoTo /D (section.7.1) >>
endobj
384 0 obj
(7.1 Grid-World mit synthetischen Bildern)
endobj
385 0 obj
<< /S /GoTo /D (subsection.7.1.1) >>
endobj
388 0 obj
(7.1.1 Machbarkeitsnachweis ohne Rauschen)
endobj
389 0 obj
<< /S /GoTo /D (subsection.7.1.2) >>
endobj
392 0 obj
(7.1.2 Bedeutung von St\366rungen im Bildformationsprozess)
endobj
393 0 obj
<< /S /GoTo /D (subsection.7.1.3) >>
endobj
396 0 obj
(7.1.3 Grid-World mit Rauschen)
endobj
397 0 obj
<< /S /GoTo /D (subsection.7.1.4) >>
endobj
400 0 obj
(7.1.4 Zusammenfassung)
endobj
401 0 obj
<< /S /GoTo /D (section.7.2) >>
endobj
404 0 obj
(7.2 Grid-World mit realer Bildformation)
endobj
405 0 obj
<< /S /GoTo /D (section.7.3) >>
endobj
408 0 obj
(7.3 Carrerabahn: Lernen an dynamischen Systemen)
endobj
409 0 obj
<< /S /GoTo /D (subsection.7.3.1) >>
endobj
412 0 obj
(7.3.1 Modellierung)
endobj
413 0 obj
<< /S /GoTo /D (subsection.7.3.2) >>
endobj
416 0 obj
(7.3.2 Versuchsablauf)
endobj
417 0 obj
<< /S /GoTo /D (subsection.7.3.3) >>
endobj
420 0 obj
(7.3.3 Merkmalsraum)
endobj
421 0 obj
<< /S /GoTo /D (subsection.7.3.4) >>
endobj
424 0 obj
(7.3.4 Ergebnisse: Lernverlauf und erlernte Strategie)
endobj
425 0 obj
<< /S /GoTo /D (subsection.7.3.5) >>
endobj
428 0 obj
(7.3.5 Diskussion der Ergebnisse)
endobj
429 0 obj
<< /S /GoTo /D (section.7.4) >>
endobj
432 0 obj
(7.4 Zusammenfassung)
endobj
433 0 obj
<< /S /GoTo /D (chapter.8) >>
endobj
436 0 obj
(8 Res\374mee)
endobj
437 0 obj
<< /S /GoTo /D (section.8.1) >>
endobj
440 0 obj
(8.1 Zusammenfassung)
endobj
441 0 obj
<< /S /GoTo /D (section.8.2) >>
endobj
444 0 obj
(8.2 Diskussion, Einordnung und Ausblick)
endobj
445 0 obj
<< /S /GoTo /D (section.8.3) >>
endobj
448 0 obj
(8.3 Fazit)
endobj
449 0 obj
<< /S /GoTo /D (appendix.A) >>
endobj
452 0 obj
(A Eindeutigkeitsbeweis des Zufallsoperators f\374r Observationen)
endobj
453 0 obj
<< /S /GoTo /D (section*.161) >>
endobj
456 0 obj
(Literatur)
endobj
457 0 obj
<< /S /GoTo /D [458 0 R /Fit ] >>
endobj
461 0 obj <<
/Length 640
/Filter /FlateDecode
>>
stream
xڅSo0~_D"ı7nch0EH0i6)8&s%hH|όl#3u/!hNV
᪢ZkRA*jM~d+,9.u6څ?W j^B`OqE4Gr3m
f_FU%T]J2aj-k~38 c&[wa? ӢR@֔F6k Y˕ڵȅʶ*\b}S.3(Ba78pO@.fp]DmEɩT#M]#ڍ\.r
-;}^[B;RgWq莩*O9yp'r@Mp'uX @?3fbBjApµ 9j)C;=DH!t8aB)K,MzB֕LE9HKAͨUndfB(Ugz`FyѧԦTuMZfoG`2 i9~1ل{h(0l|ޭ[> endobj
459 0 obj <<
/Type /XObject
/Subtype /Form
/FormType 1
/PTEX.FileName (./0_deckblatt/figures/Logo_UOS_Mitte_rot.pdf)
/PTEX.PageNumber 1
/PTEX.InfoDict 468 0 R
/BBox [0 0 391 102]
/Resources <<
/ProcSet [ /PDF ]
/XObject <<
/Fm1 469 0 R
>>>>
/Length 23
/Filter /FlateDecode
>>
stream
x+TT(c}\C|@ 1
endstream
endobj
468 0 obj
<<
/CreationDate (D:20060302120926+01'00')
/ModDate (D:20060302120926+01'00')
/Producer (Mac OS X 10.4.5 Quartz PDFContext)
>>
endobj
469 0 obj
<<
/Length 470 0 R
/Type /XObject
/Subtype /Form
/FormType 1
/BBox [ 0 0 391 102]
/Resources 471 0 R
/Filter /FlateDecode
>>
stream
xm\97yM\mmѡ#MUF;WY*AdB~^?77?8ǟ7AD$"*7ƺ8K(kB+k`/fR~gcRZZ6KLI_/*z+E)we
Vol߉:ƸnXDAuמx*r] Gȼ# 2A
NQW]/Xl"ibzLW&I_eKK$X'C̆d2blMvQ3Y8H3ffg{5Oai||>}BՅJPeOmgҎvvs
W:BR.&pqwǢ/r"pR+[8W'btE]DJ YP<
DP0:/rq\ݱ}6_8X'y>480Axjh>:?5euv}iRQ _#V]t)tGur$aUѮwsf5q
k@R
F,5ݧ\̣1Tͫ>aEU-+Q=ӔCcWbMLPI?xt Ct_M&Qp^ȢBf{MLю`UnÌlPc1N#KqzyN3P#FlE3NITI)íՈvU&&,Ñ]t: ,sla_zTq/B#LĄ-r&[ո8fM#אPU$3;-7Q&bŨJuBEv!a7֫ nFjC5o:PW^'Gvj3TI=L/:8=xsѠppwUi=%#o6j#8Q!=Yí%O>A6?>FE{#4M3 Ł-RXU'Jכ%NTQ]I;D{Rsx_қνFE;sϡP/+iZ2SW$CB~4h:;}Oӭ.+~i{Le究)y苾Fq/Yzu1ZYL)o:BXAbQ#}]VҵЈ }@7pv_;mX鎋:)J(QM'eK\svG(PZe%<Fd5PPwJ:\f 8!d*&LnA8$`TIJavp2_]zd
iHV&瘬B
{$4)lh6uAA
35R|Q?^'-̥TWRC)!
(<#k6(<5D{SH-n0e䧋j¹x^-öDnqOj6]*$Sv9a
/}EԫӚ7PBҦTv,4ѳX4ls~qOPV 1@MsvZ+IkE9G(3.OYTo*
%pWIpmW$ۇjUQv`hC#gM)*EjR=bV
U )믚g
qWEԟYu_$|~lrj>6*fC3><"{Μ3cg(v3ѿ-uoDIv:vߪj;taχMfZ͝B;xb8]f,T^]) aY~RjՊ4Z.KTC_{ymswIx7ºJ<ѐvnL[ Vp(}:D[-IVpʋkr)_UTy9-J{:]yihvdnB'4M3
IjX6*FM;R =1kAig7y46IvPӛ)*M٨AViTT1 #.O" `4Dv{Ex +=(nR繍0}mk?!P\:w#|uƪ
IyC{k[PIJaoʴTA]1[w*ޘop%"mhc!NEQcB5z:/AVmE/Zb?xYEvgIT\^6:lE쇥u@P 2ɴ
.Q#RH"dtQyD ~hH`D@[_LC=btv}}}CJ68,h?+MܓwPO
Usxs^0H&QpRC3ų̈̀AFV#iݫnoYep'5(y;q[w /99%K&@lْ[KM`TCvc CH
/F_aF\Pzo^ގ[nw.tđAT *jtZ$Mb4O}vT.x;Tg-n,QiGTf/^pGYx87#hc=%LA yb!}^d:$(2;h6W'GqCUx`TڋjMDPf[ڈobІU~Gi6Y}[b_tE`y9RmS~c>MT粒X|r%t+'ZQ䈉O#
.E&kI}aג&0 -3s̰zI_WVOb
N,'}>:c'cX-q#uTrT]T⪧c>_Ih`JثId
x>hWj)I7.^-]Eۣ)U)ZMnaRJ{E꣗O@nz˺-y"WL@^ëXV%Lb
nV(qFSlhߖӱEXYau[%YJ_8|"JLXW vNTW YES°"IS$#IPE
52%?\ir4=%
(;+/rwZ9Ytmf^QUhBQ/>zJʳD ؗHἸuN
t;zxqTz=C:UŃ9p@Gp}oJ0~'`(kp
}tDWo?! *'^D/ϐXhWB~r_?)`GE8˗k:7CA,#]Qlvh."U"#*Zv3U)+ξ=