J-j}ndZddlZddlZddlZddlmZGddeZGddeZGd d e Z y) a) VD-CMA and VkD-CMA Usage examples, VD-CMA: >>> import cma >>> from cma import restricted_gaussian_sampler as rgs >>> es = cma.CMAEvolutionStrategy(20 * [1], 1, ... rgs.GaussVDSampler.extend_cma_options({ ... 'seed': 6, ... 'ftarget': 1e-8, ... 'verbose': -9, # helpful for automatic testing ... })) >>> es = es.optimize(cma.fitness_transformations.Rotated(cma.ff.cigar, seed=6), iterations=None) >>> assert es.result.fbest <= 1e-8 >>> assert es.result.evaluations <= 6780, es.result.evaluations # was: == 6480 6372 6480 6144 It is recommended to always use `extend_cma_options()` to set the options appropriately, even when no other options are passed through. >>> len(rgs.GaussVDSampler.extend_cma_options()) 2 >>> len(rgs.GaussVkDSampler.extend_cma_options()) 3 The use case for VkD-CMA looks identical: >>> es = cma.CMAEvolutionStrategy(20 * [1], 1, ... rgs.GaussVkDSampler.extend_cma_options({ ... 'seed': 7, ... 'ftarget': 1e-8, ... 'verbose': -9, # helpful for automatic testing ... })) >>> es = es.optimize(cma.fitness_transformations.Rotated(cma.ff.cigar, seed=3), iterations=None) >>> assert es.result.fbest <= 1e-8 >>> assert es.result.evaluations < 6210, es.result.evaluations # was == 6204 TODO: correct the interface of __init__, remove unnecessaries TODO: 2017/05/10: pass the option to sampler 2017/05/10: how to give sigma to update? MEMO: 2017/05/08: line 2958 of evolution_strategy.py: cc is assigned from sp.cc 2017/05/08: line 3021 of evolution_strategy.py: `weights` are multiplied by c1 and cmu 2017/05/08: line 3021 of evolution_strategy.py: first element of `vectors` is pc 2017/05/07: hsig interface 2017/05/07: `CMAAdaptSigmaNone` not working 2017/05/07: `dimension` passed to __init__ in not int. 2017/05/06: 'AdaptSigma = CMAAdaptSigmaTPA' won't work. AssertionError happens in `_update_ps`. 2017/05/06: `correlation_matrix` is not declared in `StatisticalModelSamplerWithZeroMeanBaseClass`. However, it is used in `evolution_strategy.py`. 2017/05/06: the following line of code in `ask_geno` assumes that the result of `sample` is an ndarray, rather than list. ary = self.sigma_vec * self.sm.sample(Niid)/ N),StatisticalModelSamplerWithZeroMeanBaseClassc eZdZdZeddZejjdfdZ ddZ ddZ edZ edd Z ed Zd Zd Zd ZddZdZedZedZedZedZdZdZddZddZdZdZy)GaussVDSampleraRestricted Gaussian Sampler for VD-CMA VD-CMA: Linear Time/Space Comparison-based Natural Gradient Optimization The covariance matrix is limited as C = D * (I + v*v^t) * D, where D is a diagonal, v is a vector. Reference --------- Youhei Akimoto, Anne Auger, and Nikolaus Hansen. Comparison-Based Natural Gradient Optimization in High Dimension. In Proc. of GECCO 2014, pp. 373 -- 380 (2014) NcB|xsi}|jdtd|S)zreturn correct options to run `cma.fmin` or initialize `cma.CMAEvolutionStrategy` using the `GaussVDSampler` AKA VD-CMA-ES F) CMA_active CMA_sampler)updateroptss m/Users/jameslopez/projects/TradingBot25/.venv/lib/python3.12/site-packages/cma/restricted_gaussian_sampler.pyextend_cma_optionsz!GaussVDSampler.extend_cma_optionsJs* zr 5$24 5 Fc t||_tj|d}|jdkr t d||_||_|j|jtj|jz |_ tj|j|j|_ tj|j|_|j|jz |_|jdz|_tj"|j|_||_y#t$r*||_tj |j}YKwxYw)6pass dimension of the underlying sample space Tcopy z6Warning: Not advised to use VD-CMA for dimension < 10.N)lenNnparray TypeErroronesprintrandndvecmathsqrtvvecdotnorm_v2norm_vvnvnnzerospc_debug)self dimensionrdebugstd_vecs r __init__zGaussVDSampler.__init__Us &^DFhhyt4G 66B; J K  JJtvv&466):: vvdii3 ggdll+ ))dkk)77A:((466"  &DFggdffoG &s'E/E54E5c tjt|Dcgc],}|j|j |j .c}}|Scc}w)return list of i.i.d. samples. :param number: is the number of samples. :param update: controls a possibly lazy update of the sampler. )rasarrayrange transformrr)r*numberr iXs r samplezGaussVDSampler.samplejsJ JJ9>v GAT^^DJJtvv. / G I Hs1Ac tj|d}tj|dk\sJ|ddtjtj|ddz }|j |\}}}tj|dkDt }dtj||z } tj|ddd } tj|| d|dz} || d|} d|z |jz|tj|d |z z| zztj| | zz|_ |j|j|j\} }}}|dk(r?tj |j"}tj |j"}n8|j%|j&|j| |j(z | \}}|dk(r?tj |j"}tj |j"}nA|j%|j&|j|j|j(z \}}||z||z|zz}||z||z|zz}||zdkDr|j+|j(|j&|j|j,|j| ||||| \}}d}t/|d |j,ztjtj||z }t/|d |j(tj|z j/z}n@tj |j"}tj |j"}d}|xj0||zz c_|xj(||zz c_tj|j0|j0|_tj|j|_|j0|j,z |_|j&d z|_y) d``vectors`` is a list of samples, ``weights`` a corrsponding list of learning rates TrrNrdtype?@gffffff?r)rrallsumabs _get_paramsintr"argsortr1r(rr _alpha_avec_bsca_invavnnr&r#r'r_pvec_and_qvecr%r_ngv_ngdr$minr!)r*vectorsweightshsigwwccconecmumumueffidxsarywalphaavecbscainvavnnpvec_muqvec_mupvec_oneqvec_onepvecqvecngvngdupfactors r r zGaussVDSampler.updatets XXgD )vvbCi   VbffRVVBqrF^, ,((, D# VVBF# &bffRn$jjnTrT"zz'"3s8a<0 s3BxL7dgg%tyy G::0) "q$)00&*%B%B HHdll&$"tT7 !8hhtvv&Ghhtvv&G#22477DLL37$))3CQ H GW 19xx'Hxx'H!%!4!4TWWdll59WWtyy5H"J HhW}td{X55W}td{X55 :>}}TYY4;;%)\\5$g%)41HCH8,tyyS9I/JJLH8SDIIs ,C+H+H+J%JKH((466"C((466"CH X^#  X^# vvdii3 ii - ))dkk)77A:rcFdtjd|zz }tj|dzd|zt|z d|z zzd|zz }|dkr!dd|z t|z z dd|z zdzz }nd}d}d|dzz|z }d|d|dzzz|zz }||z }||||fS)Nr=rr?@r)rr max)r&r#gammarVbetarXrWrYs r rFz'GaussVDSampler._alpha_avec_bsca_invavnnsdiig .. '1*g S'A %K( 7], 3;3;#c(22sS7]7JQ6NNDEDUAX~$dS5!8^+s22*dD'))rctj||}t|tr:|dk(r5|dz|d|zz |||zzzz dz }||z|dzdz|zdz |zz }||fStj||dz|d|zz |||zjzjzz dz }tj|||jzjtj |dzdz|zdz |z }||fS)Nrrr=r?)rr" isinstancerDTouter)r%r#yrKy_vnr^r_s r rGzGaussVDSampler._pvec_and_qvecsvva} gs #1 a4'S7]3tq2vGG#MD!8a#  73>"DDD Tz 66'1a4'S7]*C1r6**,//+0$025$67D66'D133J>>BHHq3(C/55$56DTzrc \| |d|zz d|z| |zz|tj|| z|zz zz } | |z |tj| |zd|tj||zzz |zz } | |z ||z d|z|| zztj| ||zz zz } || z}| |fS)Nr=r?)rr")rr%r&r$r#rVrWrXrYr^r_rvecsvecr`ras r rHzGaussVDSampler._ngv_ngdsesW}- 7]tby )GbffR6F,F,L LNNd{TBFF4$99 $W-- -/1899Vmefn 7]rDy )BFF4,=,B BDDTkCxrc |jdt|jdz dz d}|jdd||jz z|jdzd|z|jz zz }|jd|dz|jd zd z|zz }|jd td |z |d z|dz d |z zz|jdzd z|zz }|||fS) Ncfactorg@g@?rNrdr?rOg?rrPr=)getrerrI)r*rRkwargsrrrNrOrPs r _get_params2zGaussVDSampler._get_params2s**YTVVb[C,?(EF ZZrEDFFN2"rEzDFF'::<=zz&'B,466C>F* *rc|j|tjd|jzdz t j ||j z|j zzz}|S);transform ``x`` as implied from the distribution parametersr=rrr r#rr"r%r*rrls r r3zGaussVDSampler.transform's[ IIdiidll(:;cARVV twwF''"" #rc||jz }|dtjd|jzz dz t j ||j z|j zz }|Srxrrs r rz GaussVDSampler.transform_inverse-sa  M cDIIcDLL011C7266 tww<''" "rctz2return inverse of associated linear transformationrr*resets r to_linear_transformation_inversez/GaussVDSampler.to_linear_transformation_inverse3rrctz'return associated linear transformationrrs r to_linear_transformationz'GaussVDSampler.to_linear_transformation7rrctzreturn scalar correction ``alpha`` such that ``X`` and ``f`` fit to ``f(x) = (x-mean) (alpha * C)**-1 (x-mean)`` rr*meanr6fs r !inverse_hessian_scalar_correctionz0GaussVDSampler.inverse_hessian_scalar_correction; "!rcV|xjtj|zc_|Sr)rrr r*factors r __imul__zGaussVDSampler.__imul__As TYYv&&  rr)T)rF)__name__ __module__ __qualname____doc__ staticmethodrrrandomrr.r7r rFrGrHrvrCrrrpropertyrrrrr3rrrrrrr rr>s )+ u*<| * *     2    6""   """ rrceZdZdZeddZejjdfdZ ddZ dZ dZ d Z d Zdd Zd Zed ZedZedZedZdZdZddZddZdZdZdZdZy)GaussVkDSampleraRestricted Gaussian Sampler for VkD-CMA O(N*k^2 + k^3) Time/Space Variant of CMA-ES with C = D * (I + V * V^T) * D References ---------- [1] Youhei Akimoto and Nikolaus Hansen. Online Model Selection for Restricted Covariance Matrix Adaptation. In Proc. of PPSN 2016, pp. 3--13 (2016) [2] Youhei Akimoto and Nikolaus Hansen. Projection-Based Restricted Covariance Matrix Adaptation for High Dimension. In Proc. of GECCO 2016, pp. 197--204 (2016) NcD|xsi}|jddtd|S)zreturn correct options to run `cma.fmin` or initialize `cma.CMAEvolutionStrategy` using the `GaussVkDSampler` AKA VkD-CMA-ES F)r AdaptSigmar )r rr s r rz"GaussVkDSampler.extend_cma_optionsTs- zr 5#($35 6 rTc  t||_tj|d}||_d|_d|_||_ |jdd|_ d|_ |jr5|jdd|_ |jd|jdz |_d|jcxkr |jcxkr|jksJJ|jd d |_|jd |j|_|jd d |_|jdd|_|jdd|_t)d|jdz |_d|jz |_d|j,z dz |_|jdd|_|jddd|jz z |_d|_d|_|jdd|_||_tj<|j|jf|_tj<|j|_ tj<|j|_!tj<|j|_"d|_#y#t$r*||_tj |j}YwxYw)rTrr=k_initrkminkmaxr k_inc_condg>@ k_dec_condk_adapt_factorg9v?factor_sigma_slopeg?factor_diag_sloperg$@r?csg333333?dsrdg@Fr:r,N)$rrrrrrrsigma sigma_fackadaptrtkk_activerrrrrrrreaccepted_slowdown k_adapt_decay k_adapt_waitrr flg_injectionpsr)Dr'VSr(dxU)r*r+rrrur-s r r.zGaussVkDSampler.__init___sC &^DFhhyt4G   Ha( ;; 61-DI 6466A:6DI8dii8$&&8 98 98$jjtd tj(|j,ztj(dtj|jnd|j|jjd|jdzzz|_|j>jutj(d|jnz|j2z |j>jx||zz |_?tj(d|jnz|_|j@|jkD}!|!|j|jkz}!|!tj d|jnd|jz|jkDz}!|!tj |jz|jkz}!|!tj tj |j||jkz}!|j|jkDd|jnd|jz|jkz}"|"|j~d|jdkz}"|j@|jkDr |!r ||_t!ttt_j|j|jz|jdz|jx|_}#tj|jjtj0|#|z |j$ff|_5tj|j$|#| zdzf|_4|j| |j\}}}d|_ n|j@||jzkDrtj|"rtj|"}$ttj|$|j}%|jj|$|_5|jnd|$jd|$|jnd|%d|jn|%d|%x|_|_|j| |j\}}}t_j`|js|j$z d z } |xj,| zc_|xjd| zc_2ycc}wcc}w)r9TrrNr:rr;r=r>opt_convrsr?)decaydim z&no first mirrored vector found for TPAz'no second mirrored vector found for TPAgGz?gGzr)axisF)PrrrrrArBr@rCrDr"rEr1rrrrrIfloatrrloglast_log_sigmar last_log_dr'last_log_cond_corrExponentialMovingAverager ema_log_sigmar ema_log_d ema_log_sitr_after_k_incrr2rrrallclose RuntimeErrorwarningswarnRuntimeWarningargmaxargminrrrexprr(r rrrjreigh_get_log_determinant_of_covr rMlnsigma_change lndiag_changelnlambda_changerrrrrrrreceilrvstackemptyany logical_not count_nonzeroshape)&r*rJrKkarrMrNrOrPrQrRrSrTrUlamr5nlistndxipimnyinner alpha_actrLrankUrVDDRidxeigrfrgL gmean_eigflg_k_increaseflg_k_decreasenewkflg_keepnew_ks& r r zGaussVkDSampler.updates ]] FF XXgD ) VbffRVVBqrF^, ,vvbCi   ((Q/ D# VVBF# &bffRn$jjnTrT"zz'"3s8a<04::= s3BxL'lQ ;;wtZ8#b%*tvv*=">>DM"$&&"4D !BFF466N2DO&(hhtvv&6D #!9mmd&<&<:?0@)@BF'')JJ466466B;?34 #:FEIIAGdNTQX%6%;q2v%FFGIE"ffSbkmmrwwtvvcr{/Ce/KLDFF1crc6N#'99T?dgg6F#GDFF1eai< S[GEIIAGdNTQX%6%;q2v%FFGIE"ffSbkmmrwwtvvcr{/Ce/KLDFF1crc6N"$''#'"2dTVVm5F5F"FDFF1bh; GaKEIIAGdNTQX%6%;q2v%FFGIE"ffSbkmmrwwtvvcr{/Ce/KLDFF1crc6N&(ggcAg&6$-9J9J&JDFF1bl? ##'99T?dgg6F#GDFF1eai< 66E>IINNtvva%i(**DFF1fuf9,=>@EBZZ^DbD)F!AF12J););)=!)LE5=5(D!$RVVB!G_a!8 8DMBfSbk?U2d:DFF3BK66$&&FUF"3Qq&"+~5FG772fSbk?3456Q FF3BK IINNtvva%i($&&FUF*;*=*=>@EBZZ^DbD)F!AF12J););)=!)LE5=5(D!$RVVB!G_a!8 8DMBfSbk?U2d:DFF3BKAvcr{N+--DFF3BK "'' U]RVVq&5&y!DFF1fuf9$55A?? 266$&&"+tvvcr{TVVCR['@A A CD D FF4;;=FLM  ) 9 ;;%   ! !!$((4::+F"G$J]J]"]^"0022dmm6:6L6L7MN"hhtzzDNN'BC b266$&&>1BFF1rvv FF7DFFOTVVGTVV_a/@1<152248OOD E!^^--t<rvv FF ffQw!9K!LLMN bffQZ043J3JJK#~~//3:>"$&&TVV"4--0A0AA$&&499,,"&&!dffWdffo"5!HII FF4&& '$*A*A A C"&& FF4%% &)?)? ?A A&&499, w $// 134//82=>  4#4#4 4.DMC $&&4+>+>">?@$&&1*M  DFTYY$(DFF1C(DEFDFXXtvvtby1}56DF"..q$&&9OT3#$D  ! !A(9(9$9 9bff? ~~n5H((2DII>EVVH%DF"ff%7hnnQ&78(CDFF6ENDFF56N%* *DFT]"..q$&&9OT3HHT==?$&&H3NO  ) 9Q *:s>F* *rc J|j}|tjtj||jd|jtj d|j d|zdz z|jd|z}||j|jzz}|S)rNr=) rrr"rrjr rrr)r*rr rls r r3zGaussVkDSampler.transforms]]  FF1dffSbkmm $ WWS466#2;& '# - /04s = = TVVdjj  rc ~||jz |jz }|jdk(r|S|tjtj|j d|j|dtj d|jd|jzz dz z|j d|jzS)Nrr=)rrrrr"rr rrs r rz!GaussVkDSampler.transform_inverses J # ==A Hrvvtvvnt}}-q1rwwsTVVNT]]%;;<DMM*,, ,rctrrrs r rz0GaussVkDSampler.to_linear_transformation_inverserrctrrrs r rz(GaussVkDSampler.to_linear_transformationrrctrrrs r rz1GaussVkDSampler.inverse_hessian_scalar_correctionrrc|xjtj|zc_|xjtj|zc_|Sr)rrr rrs r rzGaussVkDSampler.__imul__s5 dii''  $))F++ rc dtjtj|jztjtjd|jd|j zzS)Nr?r=)rrArrrrrs r rz+GaussVkDSampler._get_log_determinant_of_covsTRVVBFF466N++bff FF3// 0/22 2rctj|jtj|jz dztjd|jd|j zfS)aget the condition numbers of D**2 and (I + VV') Theoretically, the condition number of the covariance matrix can be at most the product of the return values. It might be safe to stop a run if the product of the return values reaches 1e14. Returns ------- float condition number of D float condition number of I + VV' rrN)rrerrIrrrs r get_condition_numbersz%GaussVkDSampler.get_condition_numberssMtvv/A5rvva$&&$&&/>Q7RRRrrr)rrrrrrrrrr.r7r rCrvrrrrrrrrr3rrrrrrr5rrr rrFs yy0d$KZ+&     9"""KK ,"""  2SrrceZdZdZddZdZy)rzExponential Moving Average, Variance, and SNR (Signal-to-Noise Ratio) See http://www-uxsup.csx.cam.ac.uk/~fanf2/hermes/doc/antiforgery/stats.pdf c||_tj||_tj||_| |_y)zs The latest N steps occupy approximately 86% of the information when decay = 2 / (N - 1). N)rrr'rrflg_init)r*rrflg_init_with_datas r r.z!ExponentialMovingAverage.__init__s3  ##++ rc|jr |jnd}|xj|d|z ||jz dzz|jz zz c_|xj|||jz zz c_y)Nr=rr)r8rrr)r*datumas r r zExponentialMovingAverage.update&s`--DJJR !A%$&&.1!44tvv=>> !utvv~&&rNr)rrrrr.r rrr rrs ,'rr) rrrnumpyr interfacesrrrobjectrrrr r@sG6n DEAEPMSBMS^'v'r