[DL Hacks]Deep Neuroevolution: Genetic Algorithms Are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning

DEEP LEARNING JP " [DL Papers] : : : : :: : : : " " http://deeplearning.jp/

• http://deeplearning.jp/

• a ::D .: EC:IC G C : :G CE G AF E: .: E .:GJCE F ;CE : ;CE :A: G ,: E • tn : D: :GECF C: ,: A : • :E , • , :G F F G I G :K :;; fdeoS RMV k TeoSNVwlgU CAD:G G I: G:E G I: ;CE E C E C C G : TsOLpdhm y ORivTcr PT a ub ::D

• • • • • •

• ) • w • x : 3 / : D . 0 0 / 0 3 / : 3 1 10 0 3/ 1 3 ( eihm y eca . / . A : 3 3C 3 ~ as r n pjs l u N o D t z sabk gu v _ Q

• (0 A A D • A D • E E i • e lp[ cU [ vx e qmr 1 A AD D B )3 AJ AB • A • JD • yu D C A bw x hsogn [ • D ][a • 1 • mOk a • D A E • 0 A I e R hf x S M R L A PD AD LA z (0 E B ~e t d a t R e [x A A

S a ) E ( S

• • • • • •

• ( DA E DC: A:C C: A : • CA A : C • • : • • • C:E A 0A : : a R h ( DA n vh g ce Ch Nf A : A Ch f : C ceL d S A A. : A a N h l wku Pfw h : C N N i m p Uh Syou a A Tfc h -1 t C N • ) a osrx • N b A C A C • ) a A • C:C:E l v G E C A C

T kG 7 )( o i N i n n A d n = tf 1- A 0 d 7 s -3 ,- p 7 7 7 s 7 tf d e n i

• •

• • •

[ T p 332 - [ 0 0 ( 1 (1 0 ( 1 0 0 0 0 0 1 0 0 1 = - d ns f s 0 ,1 1 0 ,1 1 ,1 0 d dt ko - 1 ( 1 ) 1 d sN ] d -45 73 k 6d = = e = e 395 823 = k = s d i

• A

• • NR o i L 2 LR ed A N Q A ADL Rh N ) h Ct 3C C N vS Q O ( A S O N G GD A R DQ C O D A G A O RNS A ON h T v r a OOTQ R c T S ARNS A O C N A T mfn

• edi :: • • : diS A d S mG •a • / / - .// / di F PE / / : mbT G di :: •T / : - : W nh / : po Sg d cl S : - di di

• https://github.com/uber-research/deep-neuroevolution
• https://github.com/tyo-yo/neural-GA

•

•

G B G 2

• •

• •

• zs RD H c • s Wl : s R n Wz • :t • e • • ik W C T • B • mp k • lpibno WP aW RDUD kxT U L R s W HP W zD T ; B ;

• /AAL 5APM A POE 1A AOE C MEODIN MA . ILAOEOE A OAM B M :M E E C /AAL 5APM 5AOR MGN B M 8AE B M AIA O 3A M E C DOOLN MSE MC N " • 0 POE OM OACEAN N 3A M E C DOOLN MSE MC N " " • DOOL AOE RRR NENO L C MEODIN DOI • 7632. 67:242 G :265 Zk fha goY_ DOOLN MSE MC N A GP OAM OE A O 8AE B M AIA O MANA M D A POE 1050:2. /2 :233 :265 cjlei " " mn OE A b MT ILPOE C CA

• https://arxiv.org/abs/1712.06567
• https://arxiv.org/abs/1703.03864
• http://www.sist.ac.jp/~kanakubo/research/evolutionary_computing/genetic_algorithms.html
• https://arxiv.org/abs/1711.01012