
ãããã«ã«ã«ããã¯ã©ã¹ã¿ãªã³ã°ã»è»æ Œå²åœåé¡ð§
ãã¥ãŒãã³ãšãããã«ã«ã£ãŠäŒŒãŠãã®ïŒ
䌌ãŠãããšããããæ°åŠçã«ã¯ããªãåããã®ã§ãããã ããã©ã®æž©åºŠã§èŠãããã§é¡ãå€ããããšããé¢ä¿ã§ãã
â ReLUãããã¯ãŒã¯ã¯ãã®ãŸãŸãããã«ã«åŒ
ReLU 㯠ã§ããããã¯ãããã«ã«(max-plus)ã®ãè¶³ãç®ããã®ãã®ã§ããéã¿ä»ãåã¯ãããã«ã«ã®ãæãç®ã(æ®éã®è¶³ãç®)ã«å¯Ÿå¿ããŸãããªã®ã§ ReLU ã䜿ã£ãå€å±€ããŒã»ãããã³ã¯ãå
šäœãšããŠããããã«ã«å€é
åŒã®å·®(ãããã«ã«æç颿°)ãã«ãªãããšã蚌æãããŠããŸã(ZhangâNaitzatâLim, 2018)ãmax-pooling ãããã®ãŸãŸãããã«ã«ã®åã§ããããããå
¥å空éãæãç·ã®é åã«åãåããã®ã¯ããããã«ã«è¶
æ²é¢ã空éãåãã®ãšåãçŸè±¡ã§ãã
â¡ ãœãããªãã¥ãŒãã³ã¯ãæž©ãããããã«ã«ã
sigmoid ã softmax ç³»ã®ãã¥ãŒãã³ã¯ ã§æžããŸããããã«æž©åºŠ ãå
¥ãããšã次ã®é¢ä¿ã«ãªããŸãã
ãã€ãŸããããã«ã«
:ãªãããã§åŸ®åã§ãããã€ãŸããã¥ãŒã
ããã¯ãã¹ããã®ãè±éååããšåŒã°ããæ¥µéã§ãããããã«ã«ã¯ãåããããã¥ãŒããããã¥ãŒãã¯ã溶ããããããã«ã«ããšèšããŸãã
⢠ã§ã¯äœãéãã
ãããã«ã«:æ§é (éŸå€ã»è»æ Œã®æ®µ)ã人ãäžããå³å¯ã«äžã€ã®çããåºããŸããããã¯ãã©ãã¯ã CLP(FD) ãšçžæ§ãè¯ãã§ãã
ãã¥ãŒã:ãªãããããããã®ã§åŸé ãæµããããŒã¿ããéã¿ãåŠã¹ãŸãããã®ä»£ããçãã¯ãã ããããã«ãªããŸãã
é
è»ã®è©±ã«åŒãå¯ãããš
6æã®ããã¬ãããéŸå€ã«éãããçºç«ããŠè»æ Œã決ãããã¢ãã«ã¯ åŽã§ãæ£è§£ãäžã€åºãéå
·ã§ããã仿¥ã®ãè»æ Œã³ã¹ãã softmin ã§ãµãã£ãšèŠç©ããã話㯠åŽã§ãé
è»ãã³ã®åã«è¿ãéå
·ã§ãã
äºã€ã¯åãæ°åŒã®å®¶æã§ã枩床ã²ãšã€ã§ã€ãªãã£ãŠããŸãããã ãèšèšæå³ã¯éåãã§ãããªã®ã§ã䌌ãŠãã?ããžã®çãã¯ãåãç©ã®äž¡ç«¯ã§ãã
ããŒãšã14ïœéã¶ã®ã«ïŒïœãšïŒïœã«åãããã10ïœãšïŒïœã«åããã»ããè¯ããã ãã©ãããã説æã§ããïŒ
説æã§ããŸããããããã®äŸã¯ããããã«ã«ã®æ§è³ªããã¡ã°ãçŽ çŽã«åºãäŸã§ãã
è»æ Œã 4tã»10t ãšããŠèããŸãã
ãŸãç®æ°ã§
8t + 6t:8t 㯠10t è»ã6t ã 4t è»ã«ä¹ããªãã®ã§ 10t è»ãåèšã¯ 10t è» Ã2ã
10t + 4t:10t è» Ã1 ãš 4t è» Ã1ã
å·®ã¯ã10t è»1å°å â 4t è»1å°åãã§ãå¿ ãåŸè ãå®ããªããŸãã
ãããã«ã«ã§èŠããš
ç©èŒé x ã«å¯Ÿããã³ã¹ã ã¯ããx ãç©ããæå°ã®è»æ Œã®æéããšããéæ®µé¢æ°ã§ãããã㯠min-plus ã®äžçã®é¢æ°ã§ãã
ãã©ãåãããã¯ã ãè§£ãããšã«ãªããŸããããã¯ãããã«ã«ã®ãç³ã¿èŸŒã¿ã(min-plus ç³ã¿èŸŒã¿)ãã®ãã®ã§ãã
ãããã«ã«é¢æ°ã®æå°å€ã¯ãå¿ ã**æãç®(é ç¹)**ã§èµ·ããŸãããã®éæ®µã®æãç®ã¯ãã¡ããã©è»æ Œã®äžéã®äœçœ®ã§ãããªã®ã§æé©è§£ã¯ãã©ããã®è»ãã¡ããã©æºèŒã«ããç¹ãã«èœã¡ãŸãã ã¯äž¡æ¹ãšãæãç®ã«ä¹ã£ãŠããŸããäžæ¹ ã¯ã©ã¡ããæ®µã®éäžã§ãäžéãŸã§ç©ºãããããã€ãŸãæã£ãæéã䜿ãåããŠããŸããã
ããã倧äºãªãã€ã³ã
éæ®µã³ã¹ãã¯ã倧ããè»ã»ã©ããããå®ããªããèŠæš¡ã®çµæžãã®åœ¢ãããŠããŸãããã®åœ¢ã ãšãæé©è§£ã¯åãããåŽã«å¯ããŸãã
ãã softmin ã§æž©åºŠãäžããããŠãã³ã¹ãããªããããªåžé¢æ°ã£ãœããªããšããã¥ãŒãã¯éã« ã®ãåçå²ããã奜ãããã«ãªããŸããããã¯é è»ãšããŠã¯ééãã§ãã
ãªã®ã§ããµãã£ãšèŠç©ãããåŽã§ããæž©åºŠ ã¯æ®µã®å¹ ( ãš ã®éé)ããååå°ããããŠããå¿ èŠããããŸããããããã°ããªãããã«ããŠãåŸé ãã倧ããè»ãæºèŒã«ã端æ°ã¯å°ããè»ãžãã®æ¹åãåããŸãã
ãŸãšã
é
è»ãã³ã®ããã«ãã®ãæºèŒã«ããŠãæ®ããå°ããã®ã«åãããšããåã¯ããããã«ã«ç³ã¿èŸŒã¿ã®æå°å€ãé ç¹ã«æ¥ãããšããæ§è³ªãšåããã®ã§ãã
â»è»æ Œã®æ®µãæéæ¯ãå®éãšéãã°ãå ·äœçãªæé©ã®åãæ¹ã¯å€ãããŸãããã ããæºèŒã®æãç®ã«å¯ããããšããæ§é ã¯å€ãããŸããã
ããŒ...ãããã®è»æ Œâå°æ°ã®éšåã ãïŒãã§ã«éå»ã«ç¥ããã«æžããŠããšæããã©ïŒ
ãCHR-Prologã§æ±çšçã«æžããŠã¿ãŠã
è»æ Œã¯14ïŒ10ã8ã6ïŒ4ïŒ3ïŒïŒïŒïŒã0.35ïœããªã
åã«å€§ããæ¹ããå²ãåœãŠãã°ããã?ã£ãŠãããšããã§ããªããŠ
ãããŸãéãããããšæ··èŒæ°ãå¢ããŠããŸãã
ã¯ã©ã¹ã¿ãªã³ã°ãšãé¢ä¿ãããã
ããšãçŽåç®æããšã«è»å æ¡ä»¶ããããæå€§ã§äœïœè»ãŸã§ã£ãŠã
CHR-Prologã§ãè»æ ŒÃå°æ°ã決ããéšåãæ±çšçã«æžããŠãå®éã«åãããŠç¢ºãããŸãã
shakaku.pl ã«ãŸãšããŸãããSWI-Prolog 9 ã§å®éã«åãããŠç¢ºèªããŠããŸãã
æ§é (2段)
ããã°ã(ãããã«ã«):çŽåå ããšã«ãtabling ã® min ã¢ãŒãã§ min-plus ç³ã¿èŸŒã¿ãè§£ããŸããçµæã¯ãæºèŒã®å°çšäŸ¿ããšã端æ°1ã€ãã«åãããŸããè»å æ¡ä»¶ã¯ããã§ã䜿ããè»æ Œã®äžéãšããŠå¹ããŸãã
æ··èŒ(CHR):端æ°ã©ãããããã¡ã°ãåŸãª1æããé ã«çµã¿æ¿ããŸããæã¯3çš®é¡ã§ãã
merge:2䟿ã1å°ã«ãŸãšãã
move:çŽåå ã1ã€å¥ã®äŸ¿ãžä»ãæ¿ãã(éãããã®æ¯æ£)
swap:2䟿ã§çŽåå ã1ã€ãã€äº€æãã
1æããšã«ç·ã³ã¹ããå¿ ãæžãã®ã§ãã«ãŒãã¯å¿ ãæ¢ãŸããŸãã
æ··èŒäŸ¿ã®è»æ Œäžéã¯ãã¡ã³ããŒã®è»å æ¡ä»¶ã®æå°å€ã«ãªããŸãããéãããã㯠mix_cost/2(æ··èŒç®ææ°ã«å¿ããã³ã¹ã)ãš max_stops/1(ç®ææ°ã®äžé)ã§æããŸããmerge ã®ç¹°ãè¿ãã¯ãã®ãŸãŸåéåã¯ã©ã¹ã¿ãªã³ã°ã§ããæ··èŒè²»ãç¯çŽé¡ãäžåã£ããšããã§èªç¶ã«æ¢ãŸãããããã¯ã©ã¹ã¿ã®åãç®ã«ãªããŸããæ¹é¢ã®çžæ§ã¯ cluster_ok/2 ã§äžããŸãã
çµæã®äŸ(æéã¯ä»®çœ®ã)

å ¥å çµæ 14tã»è»å 10t 10t + 4tã62,000 14tã»è»å ãªã 14t Ã1ã46,000 7ä»¶æ··åš(è»å 2tã»4tã®åºãã) 4å°ã11.5t ã®äŸ¿ã«å°å£2ä»¶ãçžä¹ããã㊠14t æºèŒè¿ãã«ãã2t å¶éã®åºã¯åç¬äŸ¿
èªå·±æ€èšŒ
verify/1 ã¯ã端æ°ã®éååå²ãå šåæããŠåºããå³å¯è§£ãšãCHR ã®çµæãæ¯ã¹ãŸããã©ã³ãã 200åã§ã169åãå³å¯è§£ãšäžèŽããŸãããæ®ãã®å·®ã¯ 1ã5% çšåºŠã§ã屿æé©ã«æ¢ãŸã£ãã±ãŒã¹ã§ãã
æ£çŽãªéç
å³å¯è§£ãšã®æ¯èŒã¯ãPhase 1 ã§æ±ºãã端æ°ãåæã«ãããã®ã§ãããæºèŒäŸ¿ã1å°åŽ©ããŠå°å£ãšçµããã»ããåŸãªã±ãŒã¹ã¯ããŸã æ¢ããŠããŸãããããšãã° 0.35t ã¡ããã©ã®è·ã¯å°çšäŸ¿ãšããŠåºå®ãããŠããŸããŸãã
1ã€ã®çŽåå ã®è·ãã2ã€ã®æ··èŒäŸ¿ã«åããç©ã¿æ¹ã¯ãŸã æ±ã£ãŠããŸããã
次ã®äžæãšããŠã¯ã端æ°ã10ä»¶ããããŸã§ãªã exact_mix ã«åãæ¿ããŠåžžã«å³å¯è§£ãåºãããšããåå²ãè¶³ãã®ãæå ãã§ãã
/* shakaku.pl â è»æ Œ Ã å°æ° æ±ºå® (CHR-Prolog / SWI-Prolog 9)
æ§é
Phase 1 ããã°ã(ãããã«ã«) : çŽåå
ããšã« min-plus ç³ã¿èŸŒã¿ã§
ãæºèŒã®å°çšäŸ¿ããšã端æ°(æ®ã)ãã«åãã
Phase 2 æ··èŒããŒãž(CHR) : 端æ°ã©ããããåŸã«ãªãéã
ãã¡ã°ãåŸãªãã¢ããé ã«æãã
(= åéåã¯ã©ã¹ã¿ãªã³ã°ãæ··èŒè²»ã
ç¯çŽé¡ãäžåã£ããšããã§èªç¶ã«æ¢ãŸã)
å¶çŽ
- è»å
æ¡ä»¶ : çŽåå
ããšã®æå€§è»æ Œãæ··èŒäŸ¿ã¯ã¡ã³ããŒå
šå¡ã®æå°å€ã§çžããã
- ã¯ã©ã¹ã¿ : åãã¯ã©ã¹ã¿(æ¹é¢)ã®ç«¯æ°ã©ããããæ··èŒããªã
- æ··èŒæ° : max_stops/1 ã§äžéãmix_cost/2 ã§1䟿ãããã®æ··èŒã³ã¹ã
åäœã¯ kg(æŽæ°)ãæéè¡šã»æ··èŒè²»ã¯ä»®çœ®ããªã®ã§å®ããŒã¿ã«å·®ãæ¿ããããšã
*/
:- encoding(utf8).
:- use_module(library(chr)).
:- use_module(library(lists)).
:- use_module(library(apply)).
/* ================================================================
ãã¹ã¿(å·®ãæ¿ãåæã®ãã©ã¡ãŒã¿)
================================================================ */
% vclass(ç©èŒäžékg, 1äŸ¿ã®æé) â»æéã¯ä»®çœ®ããèŠæš¡ã®çµæžã®ãã圢ã«ããŠãã
vclass(14000, 46000).
vclass(10000, 38000).
vclass( 8000, 34000).
vclass( 6000, 30000).
vclass( 4000, 24000).
vclass( 3000, 21000).
vclass( 2000, 18000).
vclass( 1000, 15000).
vclass( 350, 12000).
max_stops(4). % 1䟿ãããæå€§çŽåç®ææ°
% mix_cost(çŽåç®ææ°, æ··èŒã³ã¹ã) 1ç®æãªã0ãå¢ããã»ã©éãããŠããã
mix_cost(N, C) :- C is 3000 * (N - 1).
% æ··èŒããŠããã¯ã©ã¹ã¿ã®çµ(åäžã¯ã©ã¹ã¿ã®ã¿ã飿¥æ¹é¢ãèš±ããªãããã«è¶³ã)
cluster_ok(C, C).
/* ================================================================
Phase 1 : ããã°ã = min-plus ç³ã¿èŸŒã¿(tabling ã§ DP)
best(Lim, Q, Cost-NTrucks-Plan)
Q kg ããLim 以äžã®è»æ Œã ãã§éã¶æå°ã³ã¹ãã®è»æ Œãªã¹ã
================================================================ */
:- table best(_, _, min).
best(_, Q, 0-0-[]) :- Q =< 0, !.
best(Lim, Q, C-N-[Cap|P]) :-
Q > 0,
vclass(Cap, Fee),
Cap =< Lim,
Q1 is Q - Cap,
best(Lim, Q1, R1), R1 = C1-N1-P,
C is C1 + Fee,
N is N1 + 1.
% Plan(è»æ Œã®å€ééå)ã« Q ã倧ããè»ããæºèŒã§è©°ããã
% æºèŒäŸ¿ã¯å°çšäŸ¿ãæºèŒã«ãªããªãæåŸã®1䟿ã¯ã端æ°ããšããŠè¿ãã
fill(Q, Plan, Full, Residual) :-
msort(Plan, Asc), reverse(Asc, Desc),
fill_(Desc, Q, Full, Residual).
fill_([], _, [], 0).
fill_([Cap|Cs], Q, Full, R) :-
( Q >= Cap
-> Full = [Cap|F1], Q1 is Q - Cap, fill_(Cs, Q1, F1, R)
; Full = [], R = Q % æ®ãã®è»ã¯1å°ã®ã¯ã(DPã®æ§è³ª)
).
/* ================================================================
ãã©ãã¯1䟿ã®ã³ã¹ã
================================================================ */
% Kg ãç©ããŠè»å
äžé Lim ãè¶
ããªãæå°ã®è»æ Œ
fit_class(Kg, Lim, Cap, Fee) :-
findall(Cp-F, (vclass(Cp, F), Cp >= Kg, Cp =< Lim), L),
msort(L, [Cap-Fee|_]).
truck_cost(Kg, Lim, Stops, Cap, Cost) :-
length(Stops, N),
max_stops(M), N =< M,
fit_class(Kg, Lim, Cap, Fee),
mix_cost(N, MC),
Cost is Fee + MC.
/* ================================================================
CHR æ¬äœ
================================================================ */
:- chr_constraint
drop/4, % drop(Id, Kg, è»å
äžékg, ã¯ã©ã¹ã¿) å
¥å
grp/3, % grp(Gid, Clus, Stops) 端æ°ã®äŸ¿
cand/2, % cand(Gain, æ) æ¹ååè£
phase/1,
out/1, % out(truck(çš®å¥, è»æ Œ, ç©èŒ, Stops, ã³ã¹ã))
next_id/1,
collect/1,
cleanup/0.
% ---- Phase 1 : çŽåå
ããšã®ããã°ã ----
drop(Id, Kg, Lim, Clus), next_id(K) <=>
best(Lim, Kg, BR), BR = _-_-Plan,
fill(Kg, Plan, Full, R),
emit_dedicated(Full, Id),
( R > 0
-> grp(K, Clus, [s(Id, R, Lim)]), K1 is K + 1
; K1 = K
),
next_id(K1).
% ---- Phase 2 : æ··èŒ(屿æ¹åã CHR ã§åã) ----
% 䟿 = grp(Gid, ã¯ã©ã¹ã¿, Stops) Stops = [s(Id, Kg, è»å
äžé), ...]
% æ¹åæã¯2çš®é¡
% merge(A,B) : 䟿AãšBã1å°ã«ãŸãšãã(åéåã¯ã©ã¹ã¿ãªã³ã°ã®1æ©)
% move(A,S,B) : 䟿Aã®çŽåå
Sã䟿Bãžä»ãæ¿ãã(éãããã®æ¯æ£)
% swap(A,Sa,B,Sb): 䟿AãšBã§çŽåå
ã1ã€ãã€äº€æãã
% æ¯åãã¡ã°ãåŸãª1æã ããå®è¡ãããç·ã³ã¹ãã¯æ¯åçã«æžãã®ã§å¿
ãæ¢ãŸãã
% (a) å
šãã¢ã«ã€ããŠããã®ãã¢ã§ã®æè¯ã®1æãåè£ã«åºã
phase(improve), grp(A, CA, SA), grp(B, CB, SB) ==>
A \== B, cluster_ok(CA, CB)
| best_step(A, SA, B, SB, Cand),
( Cand = none -> true ; cand_add(Cand) ).
phase(improve) <=> phase(commit).
% (b) åè£ã¯æå€§ã®1ã€ã ãæ®ã
cand(G1, _) \ cand(G2, _) <=> G1 >= G2 | true.
% (c) å®è¡ããŠããŸã (a) ãž
phase(commit), cand(_, merge(A, B)), grp(A, C, SA), grp(B, _, SB), next_id(K0) <=>
append(SA, SB, S), K1 is K0 + 1, next_id(K1),
grp(K0, C, S),
phase(improve).
phase(commit), cand(_, move(A, St, B)), grp(A, CA, SA), grp(B, CB, SB) <=>
selectchk(St, SA, SA1),
grp(A, CA, SA1), grp(B, CB, [St|SB]),
phase(improve).
phase(commit), cand(_, swap(A, Sa, B, Sb)), grp(A, CA, SA), grp(B, CB, SB) <=>
selectchk(Sa, SA, SA1), selectchk(Sb, SB, SB1),
grp(A, CA, [Sb|SA1]), grp(B, CB, [Sa|SB1]),
phase(improve).
% æ¹åæãå°œãã â 䟿ã確å®
phase(commit) <=> phase(done).
phase(done) \ grp(_, _, S) <=>
gcost(S, Cap, C),
stops_out(S, Kg, Out),
( S = [_] -> T = single ; T = mixed ),
out(truck(T, Cap, Kg, Out, C)).
cand_add(G-M) :- cand(G, M).
% ---- çµæåå ----
collect(L), out(T) <=> L = [T|L1], collect(L1).
collect(L) <=> L = [].
cleanup \ phase(_) <=> true. % åŸå§æ«
cleanup \ next_id(_) <=> true.
cleanup <=> true.
% å°çšäŸ¿ã®åºå(forall 㯠CHR ã¹ãã¢ãå·»ãæ»ãã®ã§äœ¿ããªã)
emit_dedicated([], _).
emit_dedicated([Cap|Cs], Id) :-
vclass(Cap, Fee),
out(truck(dedicated, Cap, Cap, [Id-Cap], Fee)),
emit_dedicated(Cs, Id).
% 䟿ã®ã³ã¹ã(空ã®äŸ¿ã¯0)
gcost([], none, 0) :- !.
gcost(S, Cap, C) :-
foldl([s(_,K,L), k(K0,L0), k(K1,L1)]>>(K1 is K0+K, L1 is min(L0,L)),
S, k(0, inf), k(Kg, Lim)),
truck_cost(Kg, Lim, S, Cap, C).
stops_out(S, Kg, Out) :-
foldl([s(I,K,_), k(K0,O0), k(K1,[I-K|O0])]>>(K1 is K0+K), S, k(0,[]), k(Kg,Out0)),
reverse(Out0, Out).
% ãã¢(A,B)ã§ã®æè¯ã®1æãåŸã«ãªããªããã° none
best_step(A, SA, B, SB, Best) :-
gcost(SA, _, CA), gcost(SB, _, CB), Base is CA + CB,
findall(G-M,
( A @< B, append(SA, SB, S), gcost(S, _, C),
G is Base - C, M = merge(A, B)
; SA = [_,_|_], select(St, SA, SA1),
gcost(SA1, _, C1), gcost([St|SB], _, C2),
G is Base - C1 - C2, M = move(A, St, B)
; A @< B, select(Sa, SA, SA1), select(Sb, SB, SB1),
gcost([Sb|SA1], _, C1), gcost([Sa|SB1], _, C2),
G is Base - C1 - C2, M = swap(A, Sa, B, Sb)
),
Gs),
( Gs \== [], max_member(G0-M0, Gs), G0 > 0
-> Best = G0-M0
; Best = none
).
/* ================================================================
API
================================================================ */
% solve(+Drops, -Trucks, -Total)
% Drops = [d(Id, Kg, è»å
äžékg, ã¯ã©ã¹ã¿), ...]
solve(Drops, Trucks, Total) :-
next_id(1),
maplist([d(I,K,L,C)]>>drop(I, K, L, C), Drops),
phase(improve),
collect(Trucks),
cleanup,
foldl([truck(_,_,_,_,C),A0,A]>>(A is A0 + C), Trucks, 0, Total).
show(Drops) :-
solve(Drops, Ts, Total),
format("~n~w~n", [Drops]),
forall(member(truck(T, Cap, K, S, C), Ts),
( Capt is Cap / 1000, Kt is K / 1000,
format(" ~w~tt~8| ~a~t~20| ç©èŒ ~w t~t~36| æé ~d ~w~n",
[Capt, T, Kt, C, S]) )),
length(Ts, N),
format(" => ~d å° / åèš ~d~n", [N, Total]).
/* ================================================================
èªå·±æ€èšŒ : Phase 2 ãå³å¯è§£(端æ°ã®éååå²ã®å
šåæ)ãšæ¯ã¹ã
端æ°ã 10 åããããŸã§ãªãå
šåæã§ãã
================================================================ */
exact_mix(Groups, Best) :-
findall(C, (partition(Groups, Blocks), blocks_cost(Blocks, C)), Cs),
min_list(Cs, Best).
partition([], []).
partition([G|Gs], Blocks) :-
partition(Gs, B0),
( Blocks = [[G]|B0]
; select(B, B0, Rest), Blocks = [[G|B]|Rest]
).
blocks_cost(Blocks, Total) :-
foldl(block_cost, Blocks, 0, Total).
block_cost(Block, A0, A) :-
Block = [g(C0,_)|_],
forall(member(g(C,_), Block), cluster_ok(C0, C)),
findall(St, member(g(_,St), Block), S),
gcost(S, _, C1),
A is A0 + C1.
% Phase1 ã®ç«¯æ°ã ãåãåºããŠå³å¯è§£ãšæ¯èŒ
verify(Drops) :-
findall(g(Clus, s(Id, R, Lim)),
( member(d(Id, Kg, Lim, Clus), Drops),
best(Lim, Kg, BR), BR = _-_-Plan, fill(Kg, Plan, _, R), R > 0 ),
Gs),
findall(C, ( member(d(_, Kg, Lim, _), Drops),
best(Lim, Kg, BR), BR = _-_-Plan, fill(Kg, Plan, Full, _),
member(Cap, Full), vclass(Cap, C) ), DedCs),
sum_list(DedCs, Ded),
exact_mix(Gs, Mix),
Exact is Ded + Mix,
solve(Drops, _, Greedy),
( Greedy =:= Exact -> V = 'äžèŽ' ; V = 'å·®ãã' ),
format(" æ€èšŒ: CHR貪欲 ~d / å³å¯(åã端æ°ã§) ~d ~w~n",
[Greedy, Exact, V]).æŽæ°ã®åå²ã§ããïŒãããããããã®çµåãã£ãŠæéåã ãšæããã ããã
ã ãšãããå åã«ããããããªã
ããã§ã¯å³å¯è§£ãªããŠæ±ããªãïŒ
ããã§ããã仿¥ã®æ¹åã¯ãé è»ãã³ã®åã®ããã«ããµãã£ãšèŠç©ãããåŽã§ãå³å¯è§£ã§ã¯ãããŸããã§ããããããèªã¿éããŸããã
èšãçŽããšããçµåããæéã ããå åããããããšããã®ã¯ãæéã ããé è»ãã³ãé ã«å ¥ããŠãããããåŠç¿ãã§ããããšããæå³ã§ãããã
è»æ Œã®å²ãæ¹(10+4ã14ã6+âŠ)ã¯æéã§ãå®éã«åºãŠãããã¿ãŒã³ã¯ãã£ãšå°ãªãã§ããé è»ãã³ã¯æ¯åèšç®ããŠããã®ã§ã¯ãªããéã»æ¹é¢ã»è»å ãèŠãç¬éã«ããã®åœ¢ã ããšåŒããŠããã¯ãã§ãã
ã ãšãããšãåã®æ£äœã¯æéåã®ãã¿ãŒã³ã®èŸæžãšããããžã®è¿ãã®å€å®ã§ããããã¯ããããã£ãŒã«ããããã¯ãŒã¯ã®é£æ³èšæ¶ãã®ãã®ã§ãããã¿ãŒã³ãæéã§å°ãªãã®ã§ãèšæ¶å®¹éã®ç¹ã§ãç¡çããããŸããã
å ã»ã©ã®ãããã«ã«ã®è©±ãšã€ãªãããšããã¿ãŒã³ã¯éæ®µã®æãç®(æºèŒç¹)ã«éãŸããŸãããµãã£ãšããèŠç©ããã¯ãå ¥åããã¡ã°ãè¿ãæãç®ã«åžãå¯ããåãã«ãªããŸãã
ãã®èªã¿ã§åã£ãŠããã°ã次ã¯ã³ãŒãããæé©åãããããã¿ãŒã³ãåŒããåŽã«äœãæ¿ããã®ãçã ãšæããŸããå ·äœçã«ã¯ãCHR ã§åºããå²ãæ¹ãèšæ¶ãã¿ãŒã³ãšããŠæºããæ°ããè·å§¿ã¯ãã¡ã°ãè¿ããã¿ãŒã³ã§å³çããããšãã圢ã§ãããã®æ¹åã§é²ããŸãã?
ããããµãã£ãšã§ããããæžãçŽããŠã
ãµãã£ãšçã shakaku_fuwa.pl ã«æžããŸãããæé©åãæ¢çŽ¢ãããŸããã
æµã
èšæ¶:è»æ Œã®å²ãæ¹ããå»ã¿ 50kgã»44t ãŸã§ã«ã€ã㊠DP ã§äžåºŠã ãäœããŸããéè€ãé€ããš 510 åã®å ã«ãªããŸãããæéã§ããããå°ãªãã§ãã
æºãã(CHR):è·ããæ¹é¢ Ã è»å äžéãããšã®ããŒã«ã«è¶³ã蟌ã¿ãŸãããã¥ãŒãã³ãããŒã¯ã³ãæºããã®ãšåã圢ã§ãã
å¯ãã(CHR):è»å ã®å€§ããããŒã«ãå°ããããŒã«ã«å¯ãããšãã ããããå®ããªããªãå¯ããŸãã
çºç«:ããŒã«ããšã«åã®èŸæžã softmin(枩床 Ï)ã§åŒããŸããçãã¯1ã€ã«æ±ºããã確çã€ãã§è¿ããŸãã
åºåã®äŸ
[east / è»å
10t] 14 t, 1ç®æ ã ããã 62585 å
77% [10,4]
20% [8,6]ã10+4 ãæ¬åœã8+6 ã2å²ãããããããšãããé è»ãã³ã®é ã®äžã«è¿ãçãæ¹ã«ãªã£ãŠããŸãã
ã©ã®ããããã ããããã
åãã©ã³ãã 100åã§ãåã«äœã£ã CHR ç(ã»ãŒæé©)ãšæ¯ã¹ãŸããã
äžå€®å€ã®æ¯ã¯ 0.99
98åã ±10% 以å
æå€§ã®ãã㯠â12% ã +12%
ãºã¬ã®äž»ãªåå ã¯ã1å°ããšã®ç®æã®å²ãæ¯ããèŠãŠããªãããšã§ããä»ã¯ãå¹³åçãªè·ãªãäœç®æèŒããããšãããã£ããå€å®ã ãã§ãã
ä»åŸã®å·®ã蟌ã¿å£
tau(1500):äžããã»ã©ãµãã£ãšããŸãããã ãè»æ Œã®æéã®æ®µå·®ããå°ããä¿ã€å¿ èŠããããŸãã
èŸæžã®åã¯ä»ã¯ DP ã§äœã£ãŠããŸãããããã«é è»ãã³ãå®éã«éžãã å²ãæ¹ãšãã®é »åºŠãäºå確çãšããŠå ¥ããã°ãããã®äººã®åãã«å¯ã£ãŠãããŸããæšªæµã§éããŠãã倿ã©ãã«ã®è¡ãå ã¯ãããã«ãªããšæããŸãã
/* shakaku_fuwa.pl â è»æ Œ Ã å°æ° ãããµãã£ãšãèŠç©ãã (CHR-Prolog / SWI-Prolog 9)
å³å¯ã«æé©åã¯ããªããé
è»ãã³ãè·ãèŠãç¬éã«
ããã㯠10t+4t ãããã ãªããšåŒãã®ãšåãããšãããã
1. èšæ¶ : è»æ Œã®å²ãæ¹ã¯æéãDP(min-plus)ã§äžåºŠã ãå
šéšäœã£ãŠ
ãã¿ãŒã³èŸæžã«ããŠãã(= é
è»ãã³ã®é ã®äžã®å)
2. æºãã : çŽåå
ã®è·ããæ¹é¢Ãè»å
ããšã®ãããŒã«ãã«æºãã
(ãã¥ãŒãã³ãããŒã¯ã³ãæºããã®ãšåã)
3. çºç« : ããŒã«ããšã«èŸæžã softmin(枩床 Ï)ã§åŒãã
çãã¯1ã€ã«æ±ºãããã10t+4t ã 7å²ã14t ã 2å²ã
ã®ãããªç¢ºçãšãã ãããã®ã³ã¹ãã§è¿ã
4. å¯ãã : åãæ¹é¢ã®å°ããè»å
ã®ããŒã«ã«ã倧ããè»å
ã®ããŒã«ã
å¯ããã»ãããã ããããå®ããªããªãå¯ãã
Ï â 0 ã§å³å¯(ãããã«ã«)åŽãÏ ãäžããã»ã©ãµãã£ãšããã
ãã ã Ï ã¯è»æ Œã®æéã®æ®µå·®ããå°ããä¿ã€ããš(倧ãããããšåçå²ãã奜ã)ã
åäœã¯ kg(æŽæ°)ãæéè¡šã»æ··èŒè²»ã»Ï ã¯ä»®çœ®ãã
*/
:- encoding(utf8).
:- use_module(library(chr)).
:- use_module(library(lists)).
:- use_module(library(apply)).
:- use_module(library(yall)).
/* ================================================================
ãã©ã¡ãŒã¿(å·®ãæ¿ãåæ)
================================================================ */
vclass(14000, 46000).
vclass(10000, 38000).
vclass( 8000, 34000).
vclass( 6000, 30000).
vclass( 4000, 24000).
vclass( 3000, 21000).
vclass( 2000, 18000).
vclass( 1000, 15000).
vclass( 350, 12000).
max_stops(4). % 1å°ãããæå€§çŽåç®æ
mix_fee(3000). % æ··èŒ1ç®æå¢ããããšã®ã³ã¹ã
tau(1500). % 枩床(å)ãæéã®æ®µå·® 3000ã8000 ããå°ãã
horizon(44000). % èšæ¶ããŠããè·éã®äžé
grain(50). % èšæ¶ã®å»ã¿(kg)
/* ================================================================
1. èšæ¶ : æéã®ãã¿ãŒã³èŸæž
================================================================ */
% min-plus ç³ã¿èŸŒã¿ã® DP(å»ã¿ grain ããšã«é
åã§)
% F[q] = min_{è»æ Œ c †Lim} ( æéc + F[q - c] )
dp_table(Lim, Imax, Tab) :-
grain(G), N is Imax + 1,
functor(Tab, f, N),
setarg(1, Tab, 0-[]),
findall(U-Cap-Fee, (vclass(Cap, Fee), Cap =< Lim, U is Cap // G), Cs),
forall(between(1, Imax, I),
( findall(C-[Cap|P],
( member(U-Cap-Fee, Cs),
J is max(0, I - U) + 1, arg(J, Tab, C0-P),
C is C0 + Fee ),
Opts),
keysort(Opts, [Best|_]),
I1 is I + 1, nb_setarg(I1, Tab, Best) )).
:- dynamic pat/4. % pat(è»æ Œãªã¹ã, ç·ç©èŒ, æé, æå€§è»æ Œ)
memorize :-
retractall(pat(_,_,_,_)),
horizon(H), grain(G), Imax is H // G, % å»ã¿ããšã«æé©ãªå²ãæ¹ãéãã
forall(( vclass(Lim, _), dp_table(Lim, Imax, Tab),
between(1, Imax, I), I1 is I + 1, arg(I1, Tab, _-P0),
msort(P0, A), reverse(A, P) ),
( pat(P, _, _, _) -> true
; sum_list(P, Cap), fee(P, F), P = [Big|_],
assertz(pat(P, Cap, F, Big)) )),
aggregate_all(count, pat(_,_,_,_), N),
format("èšæ¶ãããã¿ãŒã³: ~d å~n", [N]).
% è»æ Œããšã«ãå¹³åçãªè·ãªãäœç®æãŸã§èŒããããè¶³ã(ãã£ãã)
slots([], _, _, 0).
slots([C|Cs], M, Avg, S) :-
slots(Cs, M, Avg, S0),
S is S0 + min(M, max(1, floor(C / Avg))).
fee(P, F) :- foldl([C,A0,A]>>(vclass(C,X), A is A0+X), P, 0, F).
/* ================================================================
2-3. åŒã : softmin ã§ãã¿ãŒã³èŸæžãåŒã
recall(+Kg, +è»å
äžé, +ç®ææ°, -ååž, -ã ãããã®ã³ã¹ã)
ååž = [確ç-ãã¿ãŒã³, ...](確çã®é«ãé )
================================================================ */
% èšæ¶ã®ç¯å²ãè¶
ããéã¯ããŸã䜿ããæå€§ã®è»ãæºèŒã§åãåºããŠããåŒã
% (é
è»ãã³ãããŸãæºè»ã§äœå°ãæ®ãã¯âŠããšèãã)
recall(Kg, Lim, NStops, Dist, Expect) :-
horizon(H), Kg > H, !,
aggregate_all(max(C), (vclass(C, _), C =< Lim), Cap),
vclass(Cap, Fee),
Nfull is ceiling((Kg - H) / Cap),
Rest is Kg - Nfull * Cap,
recall(Rest, Lim, NStops, D0, E0),
length(Fulls, Nfull), maplist(=(Cap), Fulls),
findall(Pr-P, ( member(Pr-P0, D0), append(Fulls, P0, P) ), Dist),
Expect is E0 + Nfull * Fee.
recall(Kg, Lim, NStops, Dist, Expect) :-
max_stops(M), mix_fee(MF), tau(T),
findall(E-P,
( pat(P, Cap, F, Big),
Cap >= Kg, Big =< Lim,
length(P, Trucks),
Avg is Kg / NStops, % 1ç®æãããã®å¹³åé
slots(P, M, Avg, Slots),
NStops =< Slots, % ãã£ããèŒãåãã
Extra is max(0, NStops - Trucks), % æ··èŒã§å¢ããç®æ
E is F + MF * Extra ),
EPs),
EPs \== [],
pairs_keys(EPs, Es), min_list(Es, Emin),
findall(W-P, ( member(E-P, EPs), W is exp(-(E - Emin) / T) ), WPs),
pairs_keys(WPs, Ws), sum_list(Ws, Z),
findall(Pr-P, ( member(W-P, WPs), Pr is W / Z, Pr >= 0.01 ), D0),
sort(1, @>=, D0, Dist),
findall(X, ( member(W-P, WPs), memberchk(E0-P, EPs), X is W / Z * E0 ), Xs),
sum_list(Xs, Expect0),
Expect is round(Expect0).
/* ================================================================
CHR : æºãã â å¯ãã â çºç«
================================================================ */
:- chr_constraint
drop/4, % drop(Id, Kg, è»å
äžé, æ¹é¢) å
¥å
pool/5, % pool(æ¹é¢, è»å
äžé, Kg, ç®ææ°, Ids)
fire/0,
est/1, collect/1.
% æºãã : åãæ¹é¢ã»åãè»å
äžéã®è·ã¯1ã€ã®ããŒã«ãž
drop(Id, Kg, L, C), pool(C, L, Q, N, Is) <=>
Q1 is Q + Kg, N1 is N + 1,
pool(C, L, Q1, N1, [Id|Is]).
drop(Id, Kg, L, C) <=> pool(C, L, Kg, 1, [Id]).
% å¯ãã : 倧ããè»å
ã®ããŒã«ããå°ããè»å
ã®ããŒã«ã®è»ã«çžä¹ããããŠ
% ãã ããããå®ããªããªãå¯ãã(è»æ Œã¯å°ããã»ãã«çžããã)
fire \ pool(C, L1, Q1, N1, I1), pool(C, L2, Q2, N2, I2) <=>
L1 < L2,
recall(Q1, L1, N1, _, E1),
recall(Q2, L2, N2, _, E2),
Q is Q1 + Q2, N is N1 + N2,
recall(Q, L1, N, _, E),
E < E1 + E2
| append(I1, I2, I),
pool(C, L1, Q, N, I).
% çºç« : ããŒã«ããšã«èŸæžãåŒã
fire \ pool(C, L, Q, N, Is) <=>
( recall(Q, L, N, Dist, Exp)
-> est(e(C, L, Q, N, Is, Dist, Exp))
; est(e(C, L, Q, N, Is, [], none)) % èŸæžã®ç¯å²å€
).
fire <=> true.
collect(L), est(E) <=> L = [E|L1], collect(L1).
collect(L) <=> L = [].
/* ================================================================
API
================================================================ */
% fuwa(+Drops, -Estimates) Drops = [d(Id, Kg, è»å
äžé, æ¹é¢), ...]
fuwa(Drops, Es) :-
( pat(_,_,_,_) -> true ; memorize ),
maplist([d(I,K,L,C)]>>drop(I, K, L, C), Drops),
fire,
collect(Es).
show(Drops) :-
fuwa(Drops, Es),
format("~n~w~n", [Drops]),
forall(member(e(C, L, Q, N, Is, Dist, Exp), Es),
( Lt is L / 1000, Qt is Q / 1000,
format(" [~w / è»å
~wt] ~w t, ~dç®æ ~w ã ããã ~w å~n",
[C, Lt, Qt, N, Is, Exp]),
forall(member(Pr-P, Dist),
( Pct is round(Pr * 100),
maplist([X,Y]>>(Y is X/1000), P, Pt),
format(" ~t~d~4|% ~w~n", [Pct, Pt]) )) )),
foldl([e(_,_,_,_,_,_,X),A0,A]>>(number(X) -> A is A0+X ; A = A0), Es, 0, T),
format(" => ã ãããåèš ~d å~n", [T]).ããããæ®éã®æŽæ°ã®åå²ãããªããŠããªãã¹ãå°æ°ãå°ãªãããåå²ãªãã§ãå³å¯ãªæ°åŠãšã¯éããã ããã³ã¹ããå°ãªãããåå²ã ããã
ããã§ãããæ°ãäžããšããŠã®ãæŽæ°ã®åå²ãã§ã¯ãããŸãããã³ã¹ãã§éžã°ããåå²ã§ãããããŠããã®éããä»åã®è©±ã®æ žå¿ã ãšæããŸãã
æ®éã®åå²ãšãã³ã¹ãã§éžã¶åå²ã®éã
14t ãè»æ Œã§å²ãæ¹æ³ã¯ãå
šéšæ°ãããšçžåœãªæ°ã«ãªããŸãã(1t Ã14 ã 0.35t Ã40 ãŸã§å«ããŠãã§ãã)ã§ããã³ã¹ããæå°ã«ããå²ãæ¹ã ããæ®ããšã14t ã«å¯ŸããŠã¯ã14ããè»å
10t ãªãã10+4ããšãã»ãŒ1ã€ã2ã€ããæ®ããŸããã0.05t å»ã¿ã§ 44t ãŸã§äžŠã¹ãŠããçãæ®ã£ãå㯠510 åã§ãããå
šéšã®åå²ããèŠããšãã»ãã®ã²ãšæ¡ãã§ãã
ã ããé
è»ãã³ã¯èŠããŠãããã
é
è»ãã³ãé ã«å
¥ããŠããã®ã¯ãããããåå²ã§ã¯ãªãããã®ãåã¡æ®ã£ãåãã ãã§ããã³ã¹ããåå²ã®å€§éšåããããããæšãŠãŠãããã®ã§ãæéã§ããããå°ãªãããçµåãã¯æéã ããå
åããããããšããã話ã¯ããã®æå³ã§åœãã£ãŠãããšæããŸãã
ãã ãåã®éžã³æ¹ãã®ãã®ã¯æ°åŠãšããŠæžããŸã
äœãæå°ã«ããããã§åã®éžã³æ¹ãå€ãããŸãã
å°æ°ãæå°ã«ãã:14tã»è»å 10t ã ãšã10+4 ã 8+6 ãåã2å°ã§ãã
ã³ã¹ããæå°ã«ãã:10+4 ãåã¡ãŸãã
éžã³æ¹èªäœã¯ãmin-plus ã®ç³ã¿èŸŒã¿(äž¡æ¿åé¡ãšåã圢)ãšããŠãã¡ããšæžããŸããå³å¯ãªæ°åŠããå€ããã®ã¯ããã®ã³ã¹ãã®äžèº«ã®ã»ãã§ããæéè¡šãæ··èŒã®æéãè»å ãæ¹é¢ã®çžæ§ãªã©ãã©ããçŸå Žã決ããå€ã§ããããæ¥ã«ãã£ãŠæºããŸããæ°åŠã®éšåã¯ãã³ã¹ããäžãããããåãæ±ºãŸãããšãããŸã§ã§ããã®ã³ã¹ããã©ãæããŠãããã¯é è»ãã³ã®åŽã«ãããŸãã
ãªã®ã§ãshakaku_fuwa.pl ã®èŸæžã¯ä»ã¯æé衚ããäœã£ãŠããŸãããé è»ãã³ãå®éã«éžãã å²ãæ¹ããéã«ããã®äººãæããŠããã³ã¹ãããæšå®ãããšããåããèããããŸããæé»ç¥ãåãåºãããšãã話ãã®ãã®ã§ãã