function [moves, vine] = solver(board, limit)
% based on "m"
% (http://www.mathworks.com/matlabcentral/contest/contests/34/submissions/64931)
%Allocate result cell
resultcell(:,3) = num2cell([-Inf*ones(10,1);0]);
%Get an index matrix for all those flipud/fliplr/rot90 cases --> not
%elegant at all, maybe incorporate this into the functions later
[nrow,ncol] = size(board);
ntot = nrow*ncol;
id = reshape(1:ntot,nrow,ncol);
%% Mike Add
limit = max(limit-100,0);
limit4soc = min(100,limit);
if limit4soc==0
[moves,vine] = alfi_soc(board,limit);
%disp('Winner is alfi_soc')
return
end
%% END Mike add
limitlimit = 15100;
if limit < 15100
[resultcell{1,1} resultcell{1,2} resultcell{1,3}] = alfi(board, limit);
resultcell{1,4} = 'alfi';
end
largestValue = max(board(:));
if largestValue > 200 || limit>10000
if (limit > 0)
[resultcell{3,1} resultcell{3,2} resultcell{3,3}] = robert(board, limit);
resultcell{3,4} = 'robert';
[movestemp vinetemp resultcell{2,3}] = robert(fliplr(board),limit); % monotonous should work good on flipping lr/ud as well
resultcell{2,4} = 'robert_fliplr';
idtemp = fliplr(id);
resultcell{2,1} = idtemp(movestemp);
resultcell{2,2} = idtemp(vinetemp);
[movestemp vinetemp resultcell{10,3}] = robert(flipud(board),limit);
resultcell{10,4} = 'robert_flipud';
idtemp = flipud(id);
resultcell{10,1} = idtemp(movestemp);
resultcell{10,2} = idtemp(vinetemp);
[resultcell{9,1} resultcell{9,2} resultcell{9,3}] = dofilter(board,limit);
resultcell{9,4} = 'dofilter';
end
doclusters=any(any(board(:,2:end)==board(:,1:end-1)))||any(any(board(1:end-1,:)==board(2:end,:))); %%%%
if doclusters && limit<numel(board)
[resultcell{4,2} resultcell{4,3}] = real_kurt(board);
resultcell{4,4} = 'real_kurt';
end
[resultcell{5,1} resultcell{5,2} resultcell{5,3}] = nick(board,limit);
resultcell{5,4} = 'nick';
[movestemp vinetemp resultcell{7,3}] = nick(rot90(board,2),limit);
resultcell{7,4} = 'nick_rot90_2';
idtemp = rot90(id,2);
resultcell{7,1} = idtemp(movestemp);
resultcell{7,2} = idtemp(vinetemp);
if max(cell2mat(resultcell([5 7],3))) < 2 * max(cell2mat(resultcell([1 2 3 4 9 10],3)))
[movestemp vinetemp resultcell{6,3}] = nick(board',limit);
resultcell{6,4} = 'nick_transpose';
idtemp = id';
resultcell{6,1} = idtemp(movestemp);
resultcell{6,2} = idtemp(vinetemp);
[movestemp vinetemp resultcell{8,3}] = nick(rot90(board,3),limit);
resultcell{8,4} = 'nick_rot90_3';
idtemp = rot90(id,3);
resultcell{8,1} = idtemp(movestemp);
resultcell{8,2} = idtemp(vinetemp);
end
end
%% Mike Add
siz = size(board);
ind_board = id;%reshape(1:prod(siz),siz);
north_neighbor = shiftnD(ind_board,1,0);
south_neighbor = shiftnD(ind_board,-1,0);
east_neighbor = shiftnD(ind_board,-2,0);
west_neighbor = shiftnD(ind_board,2,0);
neighbors = cat(3,north_neighbor,south_neighbor,east_neighbor,west_neighbor); %these are indices, not values
n_lut = reshape(neighbors,prod(siz),4);
function board = do_moves(board,moves)
% Perform moves on the board:
for abc = 1:size(moves,1)
board(moves(abc,:)) = [0 board(moves(abc,1))];
end
end
oldresultcell = resultcell;
alfi_soc_run_flag = false;
for i = 1:size(resultcell,1)
if ~isempty(resultcell{i,2})
[new_moves,resultcell{i,2},success_flag] = convert2soc(resultcell{i,2},limit4soc,board,siz,n_lut);
if success_flag
resultcell{i,1} = [resultcell{i,1};new_moves];
mod_board = do_moves(board,resultcell{i,1});
resultcell{i,3} = sum(mod_board(resultcell{i,2}));
else
if ~alfi_soc_run_flag
[alfi_soc_moves,alfi_soc_vine,alfi_soc_score] = alfi_soc(board,limit);
alfi_soc_run_flag = true;
end
resultcell{i,1} = alfi_soc_moves;
resultcell{i,2} = alfi_soc_vine;
resultcell{i,3} = alfi_soc_score;
resultcell{i,4} = 'alfi_soc';
end
end
end
%% End Mike Add
% [old_best,old_best_index] = max(cell2mat(oldresultcell(:,3)));
% disp(['Old winner was: ', oldresultcell{old_best_index,4}]);
[soc_best,best_index]=max(cell2mat(resultcell(:,3)));
% disp(['New Winner is: ', resultcell{best_index,4}])
moves = resultcell{best_index,1};
vine = resultcell{best_index,2};
if vine(1)~=1
vine(1)=1; %SOC compliance insurance!
end
% if size(moves,1) < limit
% [moves, vine] = use_more_moves(board, moves, vine, limit);
% end
end
%% nick
function [moves, vine, scr] = nick(board, limit)
moves = zeros(limit,2);
[nrow,ncol] = size(board);
ntot = nrow*ncol;
if (limit < nrow-1)
scr = -inf;
moves = [];
vine = [];
return;
end;
xg = repmat(1:ncol,nrow,1);
yg = repmat((1:nrow)',1,ncol);
pos = reshape(1:ntot,nrow,ncol);
fitloc = reshape(1:ntot,nrow,ncol);
fitloc(:,2:2:end) = flipud(fitloc(:,2:2:end));
%[s,r] = sort(board(:),'descend');
[s,r] = sort(-board(:));
s=-s;
rank = zeros(nrow,ncol);
rank(r) = 1:ntot;
remain = true(1,ntot);
nextfit = 1;
steps = 0;
cboard = board;
for i = 1:ntot
if remain(i)
[cy,cx] = find(rank==i);
tx = xg(fitloc(nextfit));
ty = yg(fitloc(nextfit));
[cboard,moves,remain,rank,steps]=oftl2(cx,tx,cy,ty,cboard,rank,steps,moves,pos,remain,ncol,nrow,limit,s,i);
if (steps > limit)
break;
end;
rank(ty,tx) = i;
nextfit = nextfit+1;
end;
end;
moves = moves(1:min(steps,limit),:);
[vine,scr] = real_kurt(cboard);
end
% %% michael
% % http://www.mathworks.com/matlabcentral/contest/contests/34/submissions/64578
% function [moves, vine] = use_more_moves(board, moves, vine, limit)
%
% m = size(board,1);
%
% for i = 1:size(moves,1)
% board(moves(i,:)) = [0 board(moves(i,1))];
% end
%
% extra_moves = limit - size(moves,1);
% % disp('Moves left over')
% larger_i = find(board >= max(board(vine)));
% larger_i = setdiff(larger_i,vine);
% if ~isempty(larger_i)
% end_row = mod(vine(end),m);
% if end_row == 0
% end_row = m;
% end
% end_col = ceil(vine(end)/m);
%
% larger_rows = mod(larger_i,m);
% larger_rows(larger_rows == 0) = m;
% larger_cols = ceil(larger_i/m);
%
% moves_req = abs(end_row - larger_rows) + abs(end_col - larger_cols) - 1;
% possible_moves = larger_i(moves_req < extra_moves);
% moves_req = moves_req(moves_req < extra_moves);
%
% while ~isempty(possible_moves) & extra_moves > 0
% [junk, sort_i] = sort(moves_req);
%
% possible_moves = possible_moves(sort_i);
%
% test_move = possible_moves(1);
% test_row = mod(test_move,m);
% if test_row == 0
% test_row = m;
% end
% test_col = ceil(test_move/m);
%
% if end_row > test_row
% move_rows = test_row:end_row;
% move_cols = zeros(size(move_rows)) + test_col;
% elseif end_row < test_row
% move_rows = test_row:-1:end_row;
% move_cols = zeros(size(move_rows)) + test_col;
% else
% move_rows = [];
% move_cols = [];
% end
%
% if end_col == test_col
% move_rows(end) = [];
% move_cols(end) = [];
% elseif end_col > test_col
% move_cols = [move_cols, test_col+1:end_col-1];
% move_rows = [move_rows, zeros(size(test_col+1:end_col-1))+end_row];
% else
% move_cols = [move_cols, test_col-1:-1:end_col+1];
% move_rows = [move_rows, zeros(size(test_col-1:-1:end_col+1))+end_row];
% end
%
% add_moves = move_cols*m + move_rows - m;
% add_moves = [add_moves(1:end-1)', add_moves(2:end)'];
%
% if any(ismember(add_moves(:),vine))
% possible_moves(1) = [];
% moves_req(1) = [];
% else
% moves = [moves; add_moves];
% extra_moves = limit - size(moves,1);
% vine = [vine, moves(end,2)];
%
% for i = 1:size(add_moves,1)
% board(add_moves(i,:)) = [0 board(add_moves(i,1))];
% end
%
% end_row = mod(vine(end),m);
% if end_row == 0
% end_row = m;
% end
% end_col = ceil(vine(end)/m);
%
% larger_i = find(board >= max(board(vine)));
% larger_i = setdiff(larger_i,vine);
% larger_rows = mod(larger_i,m);
% larger_rows(larger_rows == 0) = m;
% larger_cols = ceil(larger_i/m);
%
% moves_req = abs(end_row - larger_rows) + abs(end_col - larger_cols) - 1;
% possible_moves = larger_i(moves_req < extra_moves);
% moves_req = moves_req(moves_req < extra_moves);
% end
%
% end
% end
% end
%
function [cboard,moves,remain,rank,steps]=oftl2(cx,tx,cy,ty,cboard,rank,steps,moves,pos,remain,ncol,nrow,limit,s,i)
while (cx~=tx)||(cy~=ty)
dx = sign(tx-cx);
dy = sign(ty-cy);
if (dx~=0)&&(dy~=0)
correction = cboard(cy,cx+dx)<=cboard(cy+dy,cx);
dy = dy*(1-correction);
dx = dx*correction;
end
if (rank(cy+dy,cx+dx)>0)
vx = cx+dx;
vy = cy+dy;
if (dx==0)
[steps,moves,cboard,rank,remain,breakmarker]=oftl5(steps,limit,moves,pos,vy,vx,cboard,rank,remain,1,0,vx,ncol,cy,dy,cx,dx);
else
[steps,moves,cboard,rank,remain,breakmarker]=oftl5(steps,limit,moves,pos,vy,vx,cboard,rank,remain,0,1,vy,nrow,cy,dy,cx,dx);
end;
if breakmarker
break;
end
end;
steps = steps+1;
if (steps > limit)
break;
end;
moves(steps,1) = pos(cy,cx);
moves(steps,2) = pos(cy+dy,cx+dx);
cboard(cy+dy,cx+dx) = s(i); %cboard(cy,cx);
cboard(cy,cx) = 0;
rank(cy,cx) = 0;
cx = cx+dx;
cy = cy+dy;
end;
end
function [steps,moves,cboard,rank,remain,breakmarker]=oftl5(steps,limit,moves,pos,vy,vx,cboard,rank,remain,stepx,stepy,v_main,n_main,cy,dy,cx,dx)
if (v_main>1)&&(cboard(vy,vx)>cboard(vy-(stepy~=0),vx-(stepx~=0)))&&((v_main==n_main)||(cboard(vy+(stepy~=0),vx+(stepx~=0))>cboard(vy-(stepy~=0),vx-(stepx~=0))))
[steps,moves,cboard,rank,remain,breakmarker]=oftl3(steps,limit,moves,pos,vy,vx,cboard,rank,remain,-stepx,-stepy);
elseif (v_main < n_main)&(cboard(vy,vx)>cboard(vy+(stepy~=0),vx+(stepx~=0)))
[steps,moves,cboard,rank,remain,breakmarker]=oftl3(steps,limit,moves,pos,vy,vx,cboard,rank,remain,stepx,stepy);
else
remain(rank(cy+dy,cx+dx)) = 0;
breakmarker = 0;
end;
end
function [steps,moves,cboard,rank,remain,breakmarker]=oftl3(steps,limit,moves,pos,vy,vx,cboard,rank,remain,stepx,stepy)
steps = steps+1;
if (steps > limit)
breakmarker = 1;
else
breakmarker = 0;
moves(steps,1) = pos(vy,vx);
moves(steps,2) = pos(vy+stepy,vx+stepx);
cboard(vy+stepy,vx+stepx) = cboard(vy,vx);
cboard(vy,vx) = 0;
if (rank(vy+stepy,vx+stepx)>0)
remain(rank(vy+stepy,vx+stepx)) = 0;
end;
rank(vy+stepy,vx+stepx) = rank(vy,vx);
rank(vy,vx) = 0;
end
end
function [moves,vine,scr] = dofilter(board,limit)
[nrow,ncol] = size(board);
ntot = nrow*ncol;
moves = zeros(limit,2);
nmove = 0;
pos = reshape(1:ntot,nrow,ncol);
m1 = median(board,1);
m2 = median(board,2);
if (sum(sign(diff(m1)))>0.67*ncol)||(sum(sign(diff(m2)))>0.67*nrow)||(sum(sign(diff(m2)))<-0.67*nrow)||(sum(sign(diff(m2)))<-0.67*nrow)
d = diff(board,1,2);
mb = median(board,2);
md = median(d,2);
rmb = repmat(mb,1,ncol);
rmd = repmat(md,1,ncol);
guess = rmb+kron(md,-(ncol-1)/2:(ncol-1)/2);
bad = (abs(board-guess)>20*abs(rmd));
for i = nrow:-1:1
[nmove,moves]=oftl6(i,ncol,bad,moves,nmove,limit,pos);
if (nmove>=limit)
break;
end;
end;
moves = moves(1:min(nmove,limit),:);
for i = 1:size(moves,1)
board(moves(i,2)) = board(moves(i,1));
board(moves(i,1)) = 0;
end;
[vine,scr] = SimpleLongestPath(board);
else
scr = -inf;
moves = [];
vine = [];
end
end
function [nmove,moves]=oftl6(i,ncol,bad,moves,nmove,limit,pos)
offset = 0;
for j = ceil(ncol/2):ncol
if bad(i,j)
offset = offset+1;
else
moves(nmove+1:nmove+offset,:) = [pos(i,j:-1:j-offset+1)' pos(i,j-1:-1:j-offset)'];
nmove = nmove+offset;
if (nmove>=limit)
break;
end;
end;
end;
offset = 0;
for j = ceil(ncol/2):-1:1
if bad(i,j)
offset = offset+1;
else
moves(nmove+1:nmove+offset,:) = [pos(i,j:1:j+offset-1)' pos(i,j+1:1:j+offset)'];
nmove = nmove+offset;
if (nmove>=limit)
break;
end;
end;
end;
end
%% alfi
function [moves, vine, gain] = alfi(board, limit)
moves=[];
ab=accumarray(1+board(:),1);
cab=flipud(cumsum(flipud(ab)))-ab(end);
[m,n]=size(board);
board=[zeros(1,n+2);[zeros(m,1),board,zeros(m,1)];zeros(1,n+2)];
[m,n]=size(board);
boardab=cab(1+board);
if max(ab(2:end))>1,
% compute same-valued clusters
[BoardCluster,ClusterValue,IdxList,IdxSegments,Nclusters]=connected(board);
ClusterSize=diff(IdxSegments);
ClusterNeighb=neighb(BoardCluster,board,Nclusters);
% search between-clusters
ClustersOrder = bellman(ClusterNeighb,ClusterValue'.*sqrt(ClusterSize'));
% search within-clusters
iC=bellman_postprocess(ClustersOrder,IdxList,IdxSegments,BoardCluster);
else
% search between-pixels
% iC = bellman_pixel(board,limit,boardab);
iC = SimpleLongestPath( board );
iC = iC(end:-1:1);
iC = iC(board(iC)>0);
end
% post-processing moves
if limit>0
[board,iC,moves,limit] = postprocess(board,iC,moves,limit,boardab);
% [iC1,iC2]=ind2sub([m,n],moves);
iC1=rem(moves-1,m)+1;
iC2=(moves-iC1)/m+1;
% moves=sub2ind([m-2,n-2],iC1-1,iC2-1);
moves=iC1-1+(m-2)*(iC2-2);
end
%[iC1,iC2]=ind2sub([m,n],iC);
iC1=rem(iC-1,m)+1;
iC2=(iC-iC1)/m+1;
%vine=sub2ind([m-2,n-2],iC1-1,iC2-1);
vine=iC1-1+(m-2)*(iC2-2);
vine=vine(end:-1:1);
board=board(2:end-1,2:end-1);
gain=sum(board(vine));
end
function [board,tiC,moves,limit] = postprocess(board,tiC,moves,limit,boardab)
[m,n]=size(board);
nC=numel(tiC);
tP=board>0;
if nC>1&&limit>0, %%%
% grow laterally
d=abs(diff(tiC))==1;
E=[m*~d+d; m*d+~d];
BorderUp={repmat(1+(0:n-1)*m,[m,1]),repmat((1:m)',[1,n])};
BorderDown={repmat(m+(0:n-1)*m,[m,1]),repmat(m*(n-1)+(1:m)',[1,n])};
% tP=board>0;
tP(tiC)=false;
for n1=1:1e2,
[board,moves,tiC,tP,E,limit,ok]=postprocess_lateral(board,moves,tiC,tP,E,limit,BorderDown,BorderUp,boardab);
if ~ok, break; end
end
end
if limit>0
% grow end-points
% tP=board>0;
tP(tiC)=false;
for n1=1:1e2,
[board,moves,tiC,tP,limit,ok]=postprocess_endpoint(board,moves,tiC,tP,limit,1,m,n);
if ~ok, break; end
end
for n1=1:1e3,
[board,moves,tiC,tP,limit,ok]=postprocess_endpoint(board,moves,tiC,tP,limit,2,m,n);
if ~ok, break; end
end
end
end
function [board,moves,tiC,tP,E,limit,ok]=postprocess_lateral(board,moves,tiC,tP,E,limit,BorderDown,BorderUp,boardab)
[m,n] = size(board);
ok=0;
for ndir=1:4
offset = ndir>2.5;
coeff = (2*mod(ndir,2)-1);
K=size(tiC,2);
idx1=find(tP(tiC((1+offset):2:K-1)+coeff*E(2,(1+offset):2:K-1))&tP(tiC((2+offset):2:K)+coeff*E(2,(1+offset):2:K-1)));
for n2=numel(idx1):-1:1, %%%
switch(ndir)
case 1,
tempA=tiC(2*idx1(n2)-1)+E(2,2*idx1(n2)-1):E(2,2*idx1(n2)-1):BorderDown{1+(E(2,2*idx1(n2)-1)==m)}(tiC(2*idx1(n2)-1));
tempB=tiC(2*idx1(n2))+E(2,2*idx1(n2)-1):E(2,2*idx1(n2)-1):BorderDown{1+(E(2,2*idx1(n2)-1)==m)}(tiC(2*idx1(n2)));
case 2,
tempA=tiC(2*idx1(n2)-1)-E(2,2*idx1(n2)-1):-E(2,2*idx1(n2)-1):BorderUp{1+(E(2,2*idx1(n2)-1)==m)}(tiC(2*idx1(n2)-1));
tempB=tiC(2*idx1(n2))-E(2,2*idx1(n2)-1):-E(2,2*idx1(n2)-1):BorderUp{1+(E(2,2*idx1(n2)-1)==m)}(tiC(2*idx1(n2)));
case 3,
tempA=tiC(2*idx1(n2))+E(2,2*idx1(n2)):E(2,2*idx1(n2)):BorderDown{1+(E(2,2*idx1(n2))==m)}(tiC(2*idx1(n2)));
tempB=tiC(2*idx1(n2)+1)+E(2,2*idx1(n2)):E(2,2*idx1(n2)):BorderDown{1+(E(2,2*idx1(n2))==m)}(tiC(2*idx1(n2)+1));
case 4,
tempA=tiC(2*idx1(n2))-E(2,2*idx1(n2)):-E(2,2*idx1(n2)):BorderUp{1+(E(2,2*idx1(n2))==m)}(tiC(2*idx1(n2)));
tempB=tiC(2*idx1(n2)+1)-E(2,2*idx1(n2)):-E(2,2*idx1(n2)):BorderUp{1+(E(2,2*idx1(n2))==m)}(tiC(2*idx1(n2)+1));
end
idxZ=find(~tP(tempA),1,'first');
[nill,idxA]=max((1./(abs(board(tiC(2*idx1(n2)-1+offset))-board(tempA))+1)).*(board(tiC(2*idx1(n2)-1+offset))>=board(tempA)&board(tempA)>=board(tiC(2*idx1(n2)+offset))));
if ~nill,idxA=[]; end
if ~isempty(idxA)&&idxA<idxZ
idxZ=find(~tP(tempB),1,'first');
idxB=find(board(tempA(idxA))>=board(tempB)&board(tempB)>=board(tiC(2*idx1(n2)+offset)),1,'first');
if ~isempty(idxB)&&idxB<idxZ&&idxA+idxB-2<=limit
ok=1;
tiC=[tiC(1:2*idx1(n2)-1+offset),tiC(2*idx1(n2)-1+offset)+coeff*E(2,2*idx1(n2)-1+offset), tiC(2*idx1(n2)+offset)+coeff*E(2,2*idx1(n2)-1+offset), tiC((2*idx1(n2)+offset):end)];
E=[E(:,1:2*idx1(n2)-2),E([2,1],2*idx1(n2)-1), E([1,2],2*idx1(n2)-1), E([2,1],2*idx1(n2)-1), E(:,2*idx1(n2):end)];
tP(tiC)=false;
newmoveA=[tempA(2:idxA)',tempA(1:idxA-1)'];
newmoveB=[tempB(2:idxB)',tempB(1:idxB-1)'];
board(newmoveA(:,2))=board(tempA(idxA));
board(newmoveB(:,2))=board(tempB(idxB));
board(newmoveA(:,1))=0;
board(newmoveB(:,1))=0;
moves=[moves;flipud(newmoveA);flipud(newmoveB)];
limit=limit-size(newmoveA,1)-size(newmoveB,1);
end
end
end
end
end
function [board,moves,tiC,tP,limit,ok]=postprocess_endpoint(board,moves,tiC,tP,limit,type,m,n)
[m,n]=size(board);
ok=1;
switch(type)
case 1
mask=tP&board<2*board(tiC(1));
D=Dijkstra(mask,tiC(1),m,n);
idx1=find(mask&D-1<=limit&board>=board(tiC(1)));
if isempty(idx1),
mask=tP;
D=Dijkstra(mask,tiC(1),m,n);
idx1=find(mask&D-1<=limit&board>=board(tiC(1)));
if isempty(idx1),ok=0;end
end
if ok
if limit>150
[nill,idx2]=min(board(idx1)+1e-3*D(idx1)); %%%t3
else
[nill,idx2]=min(board(idx1).*(D(idx1)/limit).^0.2); %%%t3
end
idx0=idx1(idx2);
limit=limit-(D(idx0)-1);
for n2=D(idx0)-1:-1:1
tidx0 = idx0 + (D(idx0+1)==n2) + (~(D(idx0+1)==n2))*(-(D(idx0-1)==n2)+(~(D(idx0-1)==n2))*(m*(D(idx0+m)==n2)+(~(D(idx0+m)==n2))*(-m*(D(idx0-m)==n2))));
moves=[moves; idx0, tidx0];
board(tidx0)=board(idx0);
board(idx0)=0;
idx0=tidx0;
end
tiC=[idx0,tiC];
tP(idx0)=false;
end
case 2
mask=tP;
D=Dijkstra(mask,tiC(end),m,n);
idx1=find(mask&D-1<=limit&board<=board(tiC(end))&board>0);
if isempty(idx1),ok=0;end
if ok
if limit>300
[nill,idx2]=min(-board(idx1)+1e-3*D(idx1)); %%%t3
else
[nill,idx2]=min(-board(idx1).*(limit./D(idx1)).^0.085); %%%t3
end
idx0=idx1(idx2);
limit=limit-(D(idx0)-1);
for n2=D(idx0)-1:-1:1
tidx0 = idx0 + (D(idx0+1)==n2) + (~(D(idx0+1)==n2))*(-(D(idx0-1)==n2)+(~(D(idx0-1)==n2))*(m*(D(idx0+m)==n2)+(~(D(idx0+m)==n2))*(-m*(D(idx0-m)==n2))));
moves=[moves; idx0, tidx0];
board(tidx0)=board(idx0);
board(idx0)=0;
idx0=tidx0;
end
tiC=[tiC,idx0];
tP(idx0)=false;
end
end
end
function iC = bellman(C,D)
N=size(C,1);
c=cell(1,N); for n1=1:N,c{n1}=find(C(:,n1)); end
%C=full(C);
IDX=zeros(N,1);
cD=D;
touched=true(1,N);
while any(touched)
for touch=find(touched),
touched(touch)=false;
idx=c{touch};%find(C(:,touch));
cDnew=D(idx)+cD(touch);
idx2=find(cD(idx)<cDnew);
if ~isempty(idx2)
cD(idx(idx2))=cDnew(idx2);
touched(idx(idx2))=true;
IDX(idx(idx2))=touch;
end
end
end
[nill,idx]=max(cD);
iC=zeros(N,1);
for n1=1:N,
if idx>0
iC(n1)=idx;
idx=IDX(idx);
else break;
end
end
iC=iC(1:n1-1);
end
function iC=bellman_postprocess(ClustersOrder,IdxList,IdxSegments,BoardCluster)
[m,n]=size(BoardCluster);
iC=[];
for n1=1:numel(ClustersOrder)
idx=IdxList(IdxSegments(ClustersOrder(n1)):IdxSegments(ClustersOrder(n1)+1)-1);
if numel(idx)==1, iC=[iC,idx(1)];
else
% longest shortest-path
ThisCluster=false(m,n);
ThisCluster(idx)=true;
D=Dijkstra(ThisCluster,iC,m,n);
if n1==numel(ClustersOrder)
temp=D(ThisCluster);
else
temp=D(ThisCluster).*(BoardCluster([false(1,n);ThisCluster(1:m-1,:)])==ClustersOrder(n1+1)|BoardCluster([ThisCluster(2:m,:);false(1,n)])==ClustersOrder(n1+1)|BoardCluster([false(m,1),ThisCluster(:,1:n-1)])==ClustersOrder(n1+1)|BoardCluster([ThisCluster(:,2:n),false(m,1)])==ClustersOrder(n1+1));
end
[nill,idx0]=max(temp(:));
idx0=idx(idx0);
E=zeros(2,D(idx0)-1);
tiC=zeros(1,D(idx0));
tiC(end)=idx0;
for n2=D(idx0)-1:-1:1
if D(idx0+1)==n2, idx0=idx0+1; E(1,n2)=1;E(2,n2)=m;
elseif D(idx0-1)==n2, idx0=idx0-1; E(1,n2)=1;E(2,n2)=m;
elseif D(idx0+m)==n2, idx0=idx0+m; E(1,n2)=m;E(2,n2)=1;
elseif D(idx0-m)==n2, idx0=idx0-m; E(1,n2)=m;E(2,n2)=1;
end
tiC(n2)=idx0;
end
% now grow it
tP=ThisCluster;
tP(tiC)=false;
ok=1;
while ok
ok=0;
K=size(tiC,2);
idx1=find(tP(tiC(1:2:K-1)+E(2,1:2:K-1))&tP(tiC(2:2:K)+E(2,1:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)-1),tiC(2*idx1(n2)-1)+E(2,2*idx1(n2)-1), tiC(2*idx1(n2))+E(2,2*idx1(n2)-1), tiC(2*idx1(n2):end)];
E=[E(:,1:2*idx1(n2)-2),E([2,1],2*idx1(n2)-1), E([1,2],2*idx1(n2)-1), E([2,1],2*idx1(n2)-1), E(:,2*idx1(n2):end)];
tP(tiC)=false;
end
K=size(tiC,2);
idx1=find(tP(tiC(1:2:K-1)-E(2,1:2:K-1))&tP(tiC(2:2:K)-E(2,1:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)-1),tiC(2*idx1(n2)-1)-E(2,2*idx1(n2)-1), tiC(2*idx1(n2))-E(2,2*idx1(n2)-1), tiC(2*idx1(n2):end)];
E=[E(:,1:2*idx1(n2)-2),E([2,1],2*idx1(n2)-1), E([1,2],2*idx1(n2)-1), E([2,1],2*idx1(n2)-1), E(:,2*idx1(n2):end)];
tP(tiC)=false;
end
K=size(tiC,2);
idx1=find(tP(tiC(2:2:K-1)+E(2,2:2:K-1))&tP(tiC(3:2:K)+E(2,2:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)),tiC(2*idx1(n2))+E(2,2*idx1(n2)), tiC(2*idx1(n2)+1)+E(2,2*idx1(n2)), tiC(2*idx1(n2)+1:end)];
E=[E(:,1:2*idx1(n2)-1),E([2,1],2*idx1(n2)), E([1,2],2*idx1(n2)), E([2,1],2*idx1(n2)), E(:,2*idx1(n2)+1:end)];
tP(tiC)=false;
end
K=size(tiC,2);
idx1=find(tP(tiC(2:2:K-1)-E(2,2:2:K-1))&tP(tiC(3:2:K)-E(2,2:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)),tiC(2*idx1(n2))-E(2,2*idx1(n2)), tiC(2*idx1(n2)+1)-E(2,2*idx1(n2)), tiC(2*idx1(n2)+1:end)];
E=[E(:,1:2*idx1(n2)-1),E([2,1],2*idx1(n2)), E([1,2],2*idx1(n2)), E([2,1],2*idx1(n2)), E(:,2*idx1(n2)+1:end)];
tP(tiC)=false;
end
end
iC=[iC,tiC];
end
end
end
function B = neighb(A,V,nA)
[m,n] = size(A);
B=sparse(A(:,1:n-1),A(:,2:n),double(V(:,1:n-1)>V(:,2:n)),nA,nA)+sparse(A(:,2:n),A(:,1:n-1),double(V(:,2:n)>V(:,1:n-1)),nA,nA)+sparse(A(1:m-1,:),A(2:m,:),double(V(1:m-1,:)>V(2:m,:)),nA,nA)+sparse(A(2:m,:),A(1:m-1,:),double(V(2:m,:)>V(1:m-1,:)),nA,nA);
B=B>0;
end
function [B,C,p,r,nc] = connected(A)
[m,n] = size(A);
N = m*n ;
K = reshape (1:N, m, n) ;
V = A(:,1:n-1) == A(:,2:n);
H = A(1:m-1,:) == A(2:m,:);
G = sparse(K([V,false(m,1)]),K([false(m,1),V]),1,N,N)+sparse(K([H; false(1,n)]),K([false(1,n); H]),1,N,N);
G = G + G' + speye(N);
[p q r] = dmperm(G);
nc = numel(r) - 1;
C = A(p(r(1:nc)));
B = ones(m, n);
for a = 2:nc
B(p(r(a):r(a+1)-1)) = a;
end
end
function D = Dijkstra(A,i1,m,n)
D=inf(m,n);
D(~A)=0;
if isempty(i1),
i1=find(A,1,'first');
mode=0;
else
i1=i1(end);
mode=1;
end
D(i1)=1;
for n1=2:m*n,
idx1=D(i1+1)>n1;
idx2=D(i1-1)>n1;
idx3=D(i1+m)>n1;
idx4=D(i1-m)>n1;
%i1=unique([i1(idx1)+1,i1(idx2)-1,i1(idx3)+m,i1(idx4)-m]);
X=false(m,n);
X(i1(idx1)+1)=true;
X(i1(idx2)-1)=true;
X(i1(idx3)+m)=true;
X(i1(idx4)-m)=true;
i1=find(X);
if isempty(i1), break; end
D(i1)=n1;
end
if mode, D(D>0)=D(D>0)-1; end
end
%% kurt
function [vine, bestsom] = real_kurt(A)
moves = [];
%[bestsom,bestvine] = max(A(:));
[m,n]=size(A);
% inhoud = unique(A)';
% heur = sparse(0);
% for i=inhoud;
% heur(i+1) = sum(sum(A(A<=i)));
% end
% H=full(heur(A+1));
content = unique(A);
t=sort(A(:));
tt = cumsum(t);
%ttt = find(diff([t; 0]));
%tttt = tt(ttt);
tttt = tt(diff([t; 0])~=0);
t2 = zeros(1,size(content,1));
t2(content+1) = tttt;
H = t2(A+1);
if size(H,1) ~= size(A,1)
H=H';
end
B=A;
IND = reshape(1:m*n,[m,n]);
C = num2cell(IND);
updated = true(size(A));
iteration = 0;
while 1
G=B+H;
G=G.*updated;
[maxg,startindex] = max(G(:));
[bestsom] = max(B(:));
if maxg<=bestsom
break
end
if maxg==0
break
end
iteration = iteration+1;
i = mod(startindex,m);
j = ceil(startindex/m);
if i==0
i=m;
end
if 1<i && A(i-1,j)<=A(i,j)
[B,C,updated]=oftl4(IND,i,j,A,B,C,updated,-1,0);
end
if i<size(A,1) && A(i+1,j)<=A(i,j)
[B,C,updated]=oftl4(IND,i,j,A,B,C,updated,1,0);
end
if j<size(A,2) && A(i,j+1)<=A(i,j)
[B,C,updated]=oftl4(IND,i,j,A,B,C,updated,0,1);
end
if 1<j && A(i,j-1)<=A(i,j)
[B,C,updated]=oftl4(IND,i,j,A,B,C,updated,0,-1);
end
updated(i,j)=false;
end
[bestsom,index]=max(B(:));
vine = fliplr(C{index});
end
function [B,C,updated]=oftl4(IND,i,j,A,B,C,updated,stepi,stepj)
verplaatsindex = IND(i+stepi,j+stepj);
if all(verplaatsindex~=C{i,j})
[B(verplaatsindex),which] = max([B(verplaatsindex),B(i,j)+A(verplaatsindex)]);
if which == 2
updated(verplaatsindex)=true;
C{verplaatsindex} = [C{i,j} verplaatsindex];
end
end
end
function [vine bestScore] = SimpleLongestPath( board )
% Treats the board as a directed-acyclic-graph and finds the best
% possible solution. Upper-left wins to break cycles in the graph
% (when adjacent squares have the same value).
%
% This is basically a completely rewritten version of the sebastian()
% function, and is about 15 times faster than sebastian() on the contest
% machine.
%
% This function is only called 8 times presently on the sample set so it
% doesn't make much difference. I optimised it so extremely expecting to use it as
% a candidate-scoring/fitness function as part of some sort of larger algorithm.
%
% - Wesley
[rows cols] = size( board );
count = rows * cols;
sentinel = count + 1;
board = board(:);
score = [board; 0];
seq = reshape( 1:count, rows, cols );
north = [sentinel(ones( 1, cols )); seq(1:end-1,:)];
north = north(:);
north(board<score(north)) = sentinel;
south = [seq(2:end,:); sentinel(ones( 1, cols ))];
south = south(:);
south(board<score(south)) = sentinel;
west = [sentinel(ones( rows, 1 )); (1:count-rows)'];
west(board<score(west)) = sentinel;
east = [(rows+1:count)'; sentinel(ones( rows, 1 ))];
east(board<score(east)) = sentinel;
prev = zeros( sentinel, 1 );
[dum, order] = sort( board );
for i = 1:count
curr = order(i);
% Yes, this fixed size (4) sorting network really
% is faster than using max() for some reason.
northNeighbour = north(curr);
northScore = score(northNeighbour);
southNeighbour = south(curr);
southScore = score(southNeighbour);
if northScore >= southScore
scoreNS = northScore;
neighbourNS = northNeighbour;
else
scoreNS = southScore;
neighbourNS = southNeighbour;
end
westNeighbour = west(curr);
westScore = score(westNeighbour);
eastNeighbour = east(curr);
eastScore = score(eastNeighbour);
if westScore >= eastScore
scoreEW = westScore;
neighbourEW = westNeighbour;
else
scoreEW = eastScore;
neighbourEW = eastNeighbour;
end
if scoreEW >= scoreNS
prev(curr) = neighbourEW;
score(curr) = score(curr) + scoreEW;
else
prev(curr) = neighbourNS;
score(curr) = score(curr) + scoreNS;
end
end
[bestScore c] = max( score );
for i = 1:count
seq(i) = c;
c = prev(c);
if c == sentinel
break
end
end
vine = seq(i:-1:1);
end
% robert
function [moves, vine, score] = robert(board, limit)
% The Fragrant Honeysuckle gives up on moving and just tries to produce
% some prettier vines 8-)
% Faster to use isconnected logic, or to embed in a zeros(m+1,n+1) and guard?
% Grow a simple vine on given board
%[vine, score, dirsimple, moves] = robert2(board);
% Try constructing a boardwalk down left side
[movesb, boardb] = robert1(board, limit);
[vineb, vscoreb, dirsimple] = robert2(boardb); % Rerun vine code on modified board
% Improve score on board 47 and similar that feature a steady slope and speckle
IsMonotonous = ~(any(dirsimple(vineb) == 1) && any(dirsimple(vineb) == -1));
if IsMonotonous
% Only bother if I have a decent score already and vine goes
% solely up or solely down.
[movesm, boardm] = monotonous(board, vineb, dirsimple, limit);
[vinem, vscorem] = SimpleLongestPath(boardm); % Rerun vine code on modified board
end
% Pick best
vine = vineb;
moves = movesb;
score = vscoreb;
if IsMonotonous && vscorem > score
vine = vinem;
moves = movesm;
score = vscorem;
end
end % solver function
function [moves, board] = monotonous(boardin, vine, dir, limit)
% Focus n high-scoring vines with strong vertical orientation and high limit,
% and aim to shuffle blocking values out of high-scoring rows.
% Should never damage a unidirectional vine but ineffective unless the
% board has the structure of board 47
moves = [];
board = boardin;
[rows,cols] = size(board);
boardsize = numel(board);
lastdir = dir(vine(end)); % initialise with vine top
for i = size(vine,1)-1:-1:2 % walk down vine to polish high scores first
if abs(dir(vine(i))) == 1 % found a vertical step
valmin = board(vine(i)+dir(vine(i)));
valmax = board(vine(i)); % range of non-blocking values
% [row,col] = ind2sub([rows,cols],vine(i));
row=rem(vine(i)-1,rows)+1;
col=(vine(i)-row)/rows+1;
if lastdir > 0 && vine(i)+lastdir <= boardsize-cols
% a right edge in mid board
moveToCol = col+1;
[board,moves]=oftl(board,moves,limit,row,moveToCol,valmin,valmax,cols,rows);
row = row - 1;
if row > 0 % Needed?
moveToCol = col+1;
[board,moves]=oftl(board,moves,limit,row,moveToCol,valmin,valmax,cols,rows);
end % if second row > 0
end % if we have a midboard right edge
% Now repeat whole thing for left edges. Should be a minor pickup
% from having more moves in high-scoring areas.
% Check lastedge logic when I make 2 upward moves in a row
if lastdir == -cols && vine(i) > cols
% a left edge in mid board. Check logic.
moveToCol = col-1;
[board,moves]=oftl7(board,moves,limit,row,moveToCol,valmin,valmax,rows);
end % if we have a midboard left edge
end % vertical step
lastdir = dir(vine(i));
end % for each leaf of vine
% Have now modified board, hopefully improved it in a few high-scoring cases
end % function monotonous
function [board,moves]=oftl7(board,moves,limit,row,moveToCol,valmin,valmax,rows)
while size(moves,1) < limit % loop through all blockages (if any)
while ( board(row,moveToCol) >= valmin && ...
board(row,moveToCol) <= valmax ) && moveToCol > 1
moveToCol = moveToCol - 1;
end
if moveToCol > 1 % Found a blockage I'd like to replace
moveFromCol = moveToCol-1;
while moveFromCol > 0 && ...
( board(row,moveFromCol) < valmin || ...
board(row,moveFromCol) > valmax )
moveFromCol = moveFromCol - 1;
end
if moveFromCol > 0
% Something to replace it with exists, so do slide
for j = moveFromCol:moveToCol-1
if size(moves,1) < limit
moves = [moves; [row+rows*(j-1), row+rows*j]];
board(row,j+1) = board(row,j);
board(row,j) = 0;
else
break; % Used up all moves
end
end
else
break; % no more nonblocks to use so row finished
end
else
break; % no blocks so row finished
end % (needed in case block is in other row...?)
end % First row while loop. Break out when complete and apply
% same logic to second row
end
function [board,moves]=oftl(board,moves,limit,row,moveToCol,valmin,valmax,cols,rows)
while size(moves,1) < limit % loop through all blockages (if any)
while ( board(row,moveToCol) >= valmin && ...
board(row,moveToCol) <= valmax ) && moveToCol <= cols-1
moveToCol = moveToCol + 1;
end
if moveToCol < cols % Found a blockage I'd like to replace
moveFromCol = moveToCol+1;
while moveFromCol <= cols && ...
( board(row,moveFromCol) < valmin || ...
board(row,moveFromCol) > valmax )
moveFromCol = moveFromCol + 1;
end
if moveFromCol <= cols
% Something to replace it with exists, so do slide
for j = moveToCol:moveFromCol-1
if size(moves,1) < limit
moves = [moves; [(row+rows*j), (row+rows*(j-1))]];
board(row,j) = board(row,j+1);
board(row,j+1) = 0;
else
break; % Used up all moves
end
end
else
break; % no more nonblocks to use so row finished
end
else
break; % no blocks so row finished
end % (needed in case block is in other row...?)
end
end
function [moves, board] = robert1(boardin, limit)
% Focus n high-score boards like 8 where there is a high limit but no structure.
% Aim to construct a path from scratch. To keep the movement choices simple do
% this along an edge; a spiral in centre would require far fewer moves on
% average but would require more complex pathfinding
% Minor bug with equal values, which may upset vine constructor so add a small
% amount of noise.
moves = zeros(limit,2); mv=0;
[rows,cols] = size(boardin);
boardsize = numel(boardin);
board = boardin - reshape(1:boardsize,size(boardin))*1e-4;
Target = 1; % Pick a corner, any corner
while mv < limit % Find another block to move
[Biggest,BlockAt] = max(board(:)); % Move biggest
if Biggest <= 0 % Prob not needed
break;
end
Vt=rem(Target-1,rows)+1;
Ut=(Target-Vt)/rows+1;
% First move onto a channel that will be cleared of blocks
Vb=rem(BlockAt-1,rows)+1;
Ub=(BlockAt-Vb)/rows+1;
if Ub-Ut > 1 % Only if not next to target column
Nudge = 1-mod(Vb-1,3); % [+1,0,-1,...]
if (Nudge ~= 0) && ~(Vb==rows && Nudge==1)
mv=mv+1; moves(mv,1:2) = [BlockAt,BlockAt+Nudge];
board(BlockAt+Nudge) = board(BlockAt);
board(BlockAt) = 0;
BlockAt = BlockAt+Nudge;
end
end
% Now move towards target
Vb = rem(BlockAt-1,rows)+1; U = Ut-(BlockAt-Vb)/rows-1;
V = Vt-Vb; % Horiz and Vert moves required
while (U~=0 || V~=0) && mv<limit
if U<-1 || V==0
mv=mv+1; moves(mv,1:2) = [BlockAt,BlockAt-rows];
board(BlockAt-rows) = board(BlockAt);
board(BlockAt) = 0;
BlockAt = BlockAt-rows;
U = U+1;
elseif V>0 % Just ignore erases until I get it working
mv=mv+1; moves(mv,1:2) = [BlockAt,BlockAt+1];
board(BlockAt+1) = board(BlockAt);
board(BlockAt) = 0;
BlockAt = BlockAt+1;
V=V-1;
else
mv=mv+1; moves(mv,1:2) = [BlockAt,BlockAt-1];
board(BlockAt-1) = board(BlockAt);
board(BlockAt) = 0;
BlockAt = BlockAt-1;
V=V+1;
end
end % Sliding this block
board(Target) = -board(Target); % Ignore moved blocks
Target=Target+(mod(Ut,2))*(Vt<rows)-(~mod(Ut,2))*(Vt>1)+rows*(mod(Ut,2))*(~(Vt<rows))+rows*(~mod(Ut,2))*(~(Vt>1));
end % Finding blocks to slide
board = abs(board);
moves=moves(1:mv,:);
end % function boardwalk
function [vine, vscore, dir, moves] = robert2(board)
% Test each square on the board from low to high values to find whether vine
% can grow upwards onto an adjacent target square.
% To handle plateaux I superimpose a faint watermark on flat areas to nudge vine
% towards a space-covering pattern. On each plateau I track 4 orientations
% separately, but whenever looking up to a higher level I can pick the best.
% Quite pretty, but as all plateaux problems in testset are low-scoring I don't
% think it makes much difference.
%
dir = zeros(size(board)); % points way down vine, 0 at root
% done = false(size(board)); % have already examined this one
[rows,cols] = size(board);
boardsize = numel(board);
watermark = reshape(1:boardsize,size(board)) * 1e-6;
watermark(:,2:2:cols) = flipud(watermark(:,2:2:cols));
board = board + watermark .* (mod(board,2)-0.5);
% superimpose a faint watermark on flat areas. Direction should change
% between adjacent levels (but will fail if steps are even)
score = board; % score for a vine starting and ending here
[val, ind] = sort(board(:));
% Loop through board from low to high values to find the score associated
% with the best vine growing down from each square
% Needs more care on equal values, for now just weave up and down
% -- Could try 4 weaves and pick best, or proper handling of plateaux
for i = 1:boardsize;
test = ind(i);
stepvec = [-rows,-1,1,rows];
for step = stepvec;
targ = test + step;
if targ > 0 && targ <= boardsize && ...
(abs(step)==rows || mod(min(test,targ), rows))
% consider only adjacent squares on board
if (board(targ) > board(test)) && (board(targ)+score(test) > score(targ))
% Targ can grow from me (and has nothing better yet)
score(targ) = score(test) + board(targ);
% add test vine to targ score
dir(targ) = -step;
% and note where it came from
end
end
end
end
% Assemble vine from top down
[dum,leaf] = max(score(:));
vine = [];
while leaf
vine = [vine;leaf];
leaf = (dir(leaf)~=0)*(leaf+dir(leaf));
end
vine = flipud(vine); % return it in correct order
vscore = score(vine(end));
moves = [];
end % growvine
%% Mike Add convert2soc
function [new_moves,new_vine,success_flag] = convert2soc(vine,limit4soc,board,siz,n_lut)
% If the vine already contains 1, then just truncate the eariler part of
% the vine
if any(vine==1)
new_vine = vine(find(vine==1):end);
new_moves = zeros(0,2);
success_flag = true;
return
end
if nargin<4 || isempty(siz) || isempty(n_lut)
siz = size(board);
ind_board = reshape(1:prod(siz),siz);
north_neighbor = shiftnD(ind_board,1,0);
south_neighbor = shiftnD(ind_board,-1,0);
east_neighbor = shiftnD(ind_board,-2,0);
west_neighbor = shiftnD(ind_board,2,0);
neighbors = cat(3,north_neighbor,south_neighbor,east_neighbor,west_neighbor); %these are indices, not values
n_lut = reshape(neighbors,prod(siz),4);
end
[l2v,grown] = paths2vine(vine,siz,n_lut,limit4soc);
% We want the path which reaches the vine closest to the base without going
% over the limit of moves
attach_point = find(l2v<limit4soc-1,1,'first');
if isempty(attach_point)
% Can't connect to vine, need emergency backup plan
% disp('CANT REACH')
new_moves = zeros(0,2);
new_vine = [1]; % to conform with minimal rules
success_flag = false;
return
end
trunc_vine = vine(attach_point:end);
graft_length = l2v(attach_point);
graft_path = make_graft(vine(attach_point),grown,graft_length,n_lut);
graft_moves = [graft_path(end:-1:2),graft_path(end-1:-1:1)];
new_moves = [graft_moves(1,:); graft_moves];
new_vine = [graft_path(:); trunc_vine(:)]';
success_flag = true;
end
% idea here is to create a corridor of zeros from square one to the end of
% the already-optimized vine. If limit is 0, then just
% SOC ends 4pm PT TUESDAY
function [graft_path] = make_graft(attach_point_ind,grown,graft_length,n_lut)
graft_path = zeros(graft_length,1);
pathend = attach_point_ind;
level = graft_length+1;
for j = graft_length:-1:1
%while graft_path(end)~=1
local_nodes = nb_nodes(pathend,n_lut);
for i = 1:length(local_nodes)
if grown(local_nodes(i))>0 && grown(local_nodes(i))<level
pathend = local_nodes(i);
graft_path(j) = pathend;
level = grown(pathend);
break
end
end
end
end
function [l2v,grown] = paths2vine(vine,siz,n_lut,limit)
vine_mask = false(siz);
vine_mask(vine) = true;
grown = zeros(siz);
grown(1)=1;
level = 1;
newlygrown = true(siz);
while any(newlygrown(:)) && ~(level >= limit-2)
level = level+1;
% grow by shifting current level
shifts = shiftnD(grown,-1,0) + shiftnD(grown,+1,0) + shiftnD(grown,-2,0) + shiftnD(grown,2,0);
shifts(vine_mask) = 0;
newlygrown = shifts>0 & grown==0;
grown(newlygrown) = level;
end
grown2 = grown;
grown2(grown2==0)=Inf;
l2v = zeros(length(vine),1);
for i = 1:length(vine)
adj_vals = nb_vals(vine(i),grown2,n_lut);
l2v(i) = min(adj_vals);
end
end
% nb_nodes
function local_nodes = nb_nodes(node,n_lut)
local_nodes = n_lut(node,:);
local_nodes(local_nodes==0)=[] ;
end
function local_nb_vals = nb_vals(node,board,n_lut)
nb_nodes = n_lut(node,:);
nb_nodes(nb_nodes==0) = [];
local_nb_vals = board(nb_nodes);
end
function shifted_array = shiftnD(array, direction,padding)
% shifted_array = shiftnD(array, direction)
% This function should take the given array and shift it in the direction
% given. The direction should be the number of the dimension in which to
% shift, positive if shifting in the direction of increasing indices,
% negative if shifting in the direction of decreasing indices.
if direction ==0 %don't shift
shifted_array = array;
return
end
dims = size(array);
shiftdim = abs(direction); % this is the dimension to shift in
shiftdir = sign(direction); % this is the direction to shift in that dimension
if abs(direction)>length(dims)
%error('Direction chosen is inappropriate')
end
% construct extended matrix
for i = 1:length(dims)
c{i} = 1:dims(i);
end
c2 = c;
if shiftdir>0
c{shiftdim} = 2:dims(shiftdim);
c2{shiftdim} = 1:dims(shiftdim)-1;
elseif shiftdir<0
c{shiftdim} = 1:dims(shiftdim)-1;
c2{shiftdim} = 2:dims(shiftdim);
end
shifted_array = repmat(padding,dims);
shifted_array(c{:}) = array(c2{:});
end
%% %%%%%%% SOC FUNCTIONS %%%%%%%%%%%%%%%%%%%%%%
% I need a function for the case where a graft can't reach the vine.
% Perhaps the most likely case for this is when there are no moves allowed
% SO, we need a solver which grows the best possible vine from square onefunction [moves, vine, gain] = alfi(board, limit)
function [moves, vine, gain] = alfi_soc(board, limit)
moves=[];
ab=accumarray(1+board(:),1); % ab(i) counts the number of board locations
% with value i-1 (because of the +1)
cab=flipud(cumsum(flipud(ab)))-ab(end); % computes the reverse cumulative sum of
% the ab list, reverses it, and subtracts the
% number of times the highest value in board
% occurs (I wonder if it should be the number
% of times 0 occurs (ab(1)) instead, but I
% don't understand why this is here, so I'm
% not sure
[m,n]=size(board);
board=[zeros(1,n+2);[zeros(m,1),board,zeros(m,1)];zeros(1,n+2)]; % has the
% result of padding the board by a ring of zeros all the way around
[m,n]=size(board);
boardab=cab(1+board); %
if max(ab(2:end))>1,
% If any of the ab vector is greater than 1, then there might be
% clusters, since the same value occurs more than once in the board. We
% don't care about ab(1) because those are zero-weight nodes
% compute same-valued clusters
[BoardCluster,ClusterValue,IdxList,IdxSegments,Nclusters]=connected(board);
% BoardCluster is a label matrix of connected, identically-valued board
% regions. ClusterValue is a vector of length Nclusters such that
% ClusterValue(i) gives you the board value of one node in the region
% described by BoardCluster==i. IdxList and IdxSegments are p and r
% decribed in the comments to the connected function below. Nclusters
% is the number of clusters found. It is important to note that
% nodes with no identically-valued neighbors are just treated as
% clusters of size 1 throughout this section, so there are no nodes not
% in a "cluster"
ClusterSize=diff(IdxSegments); % vector of the number of nodes in each cluster
ClusterNeighb=neighb_soc(BoardCluster,board,Nclusters); % Finds sparse matrix
% describing which clusters neighbor one another and can be moved to, i.e.
% ClusterNeighb(i,j) = 1, iff BoardCluster==i neighbors
% BoardCluster==j and ClusterValue(i)>ClusterValue(j)
% (I'm not 100% sure on this, but it's something like that)
% search between-clusters
ClustersOrder = bellman_soc(ClusterNeighb,ClusterValue'.*sqrt(ClusterSize'));
% search within-clusters
iC=bellman_postprocess_soc(ClustersOrder,IdxList,IdxSegments,BoardCluster);
else
% search between-pixels
iC = bellman_pixel_soc(board,limit,boardab);
end
% post-processing moves
if limit>0
[board,iC,moves,limit] = postprocess_soc(board,iC,moves,limit,boardab);
[iC1,iC2]=ind2sub([m,n],moves);
moves=sub2ind([m-2,n-2],iC1-1,iC2-1);
end
[iC1,iC2]=ind2sub([m,n],iC);
vine=sub2ind([m-2,n-2],iC1-1,iC2-1);
vine=vine(end:-1:1);
board=board(2:end-1,2:end-1);
gain=sum(board(vine));
end
%% postprocess_soc
function [board,tiC,moves,limit] = postprocess_soc(board,tiC,moves,limit,boardab)
[m,n]=size(board);
nC=numel(tiC);
if nC>1&&limit>0, %%%
% grow laterally
d=abs(diff(tiC))==1;
E=[m*~d+d; m*d+~d];
BorderUp={repmat(1+(0:n-1)*m,[m,1]),repmat((1:m)',[1,n])};
BorderDown={repmat(m+(0:n-1)*m,[m,1]),repmat(m*(n-1)+(1:m)',[1,n])};
tP=board>0;
tP(tiC)=false;
for n1=1:1e3,
[board,moves,tiC,tP,E,limit,ok]=postprocess_lateral(board,moves,tiC,tP,E,limit,BorderDown,BorderUp,boardab);
if ~ok, break; end
end
end
if limit>0
% grow end-points
tP=board>0;
tP(tiC)=false;
for n1=1:1e3,
[board,moves,tiC,tP,limit,ok]=postprocess_endpoint(board,moves,tiC,tP,limit,1);
if ~ok, break; end
end
% for n1=1:1e3,
% [board,moves,tiC,tP,limit,ok]=postprocess_endpoint(board,moves,tiC,tP,limit,2);
% if ~ok, break; end
% end
end
end
%% bellman_pixel
function iC = bellman_pixel_soc(board,limit,boardab)
[m,n]=size(board);
cD=board;
IDX=zeros(m,n);
touched = false(m,n);
touched(2,2) = true;
%touched=board>0;
%valid=touched;
valid = board>0;
while any(touched(:))
touch=find(touched);
touched(touch)=false;
for dx=[1,-1,m,-m],
cDnew=board(touch+dx)+cD(touch); %%%
idx=find(board(touch+dx)>board(touch)&cDnew>cD(touch+dx));
touched(touch(idx)+dx)=valid(touch(idx)+dx);
cD(touch(idx)+dx)=cDnew(idx);
IDX(touch(idx)+dx)=touch(idx);
end
end
[nill,idx]=max(cD(:)-board(:));
if idx==1;
idx=m+2; % want idx = 1 in original board
end
iC=zeros(1,m*n);
for n1=1:m*n,
if idx>0
iC(n1)=idx;
idx=IDX(idx);
else break;
end
end
iC=iC(1:n1-1);
end
%% bellman
function iC = bellman_soc(C,D)
% C is sparse cluster neighbor matrix ClusterNeighb, and D is a weighted
% value for each cluster (single node value x sqrt(number of elements in cluster)
N=size(C,1); % Number of clusters
c=cell(1,N);
% Construct list of neighbors for each cluster
for n1=1:N,
% c{n1}=find(C(:,n1)); % c{i} = vector of cluster numbers which are reachable from cluster i
c{n1}=find(C(n1,:)); % c{i} = vector of cluster numbers which are reachable from cluster i
end
%C=full(C);
IDX=zeros(N,1);
cD=D; % This is a weight to be gained by visiting a cluster
touched = false(1,N);
touched(2) = true;
%touched=true(1,N); % start search from every cluster
while any(touched)
for touch=find(touched), % loop over clusters on the frontier
touched(touch)=false; % mark the current cluster as visited
idx=c{touch};%find(C(:,touch)); % look up the visitable clusters for the current cluster
cDnew=D(idx)+cD(touch); % This adds the weight of visitable clusters to the stored weight of the current cluster
idx2=find(cD(idx)<cDnew); % This finds which scores are better than the current score
if ~isempty(idx2)
cD(idx(idx2))=cDnew(idx2); % For neighboring clusters where the new score is better than the old, update the stored score
touched(idx(idx2))=true; % furthermore, you want to start working from these clusters again, since you have a better score you want to propagate
IDX(idx(idx2))=touch; %
end
end
end
[nill,idx]=max(cD-D);
if idx==1
idx=2; % can't use boarder padding, must use cluster 2
end
iC=zeros(N,1);
for n1=1:N,
if idx>0
iC(n1)=idx;
idx=IDX(idx);
else break;
end
end
iC=iC(1:n1-1);
end
%% bellman_postprocess
function iC=bellman_postprocess_soc(ClustersOrder,IdxList,IdxSegments,BoardCluster)
% This is the intra-cluster vine-growing algorithm
[m,n]=size(BoardCluster);
iC=[];
for n1=1:numel(ClustersOrder)
idx=IdxList(IdxSegments(ClustersOrder(n1)):IdxSegments(ClustersOrder(n1)+1)-1);
if numel(idx)==1, iC=[iC,idx(1)]; % if the current cluster is just one pixel, just add it to the list, no further path determinations needed
else
% longest shortest-path
ThisCluster=false(m,n);
ThisCluster(idx)=true; % Makes ThisCluster a logical map of the current cluster
D=Dijkstra(ThisCluster,iC,m,n); % Grows a map of number of moves to get from the
% end of the current path (iC) to anywhere within the
% cluster
if n1==numel(ClustersOrder)
% temp=D(ThisCluster); % if this is the last cluster, then we don't need to find a particular endpoint, the furthest place is fine...
temp = false(size(D(ThisCluster)));
temp(1) = true;
else
temp=D(ThisCluster).*(BoardCluster([false(1,n);ThisCluster(1:m-1,:)])==ClustersOrder(n1+1)...
|BoardCluster([ThisCluster(2:m,:);false(1,n)])==ClustersOrder(n1+1)...
|BoardCluster([false(m,1),ThisCluster(:,1:n-1)])==ClustersOrder(n1+1)...
|BoardCluster([ThisCluster(:,2:n),false(m,1)])==ClustersOrder(n1+1));
% this finds where the path across the cluster needs to exit to
% continue the path
end
[nill,idx0]=max(temp(:));
idx0=idx(idx0);
E=zeros(2,D(idx0)-1);
tiC=zeros(1,D(idx0));
tiC(end)=idx0;
% This for loop constructs a shortest path across the cluster from the
% identified start to the identified end
for n2=D(idx0)-1:-1:1
if D(idx0+1)==n2, idx0=idx0+1; E(1,n2)=1;E(2,n2)=m;
elseif D(idx0-1)==n2, idx0=idx0-1; E(1,n2)=1;E(2,n2)=m;
elseif D(idx0+m)==n2, idx0=idx0+m; E(1,n2)=m;E(2,n2)=1;
elseif D(idx0-m)==n2, idx0=idx0-m; E(1,n2)=m;E(2,n2)=1;
end
tiC(n2)=idx0;
end
% now grow it
tP=ThisCluster;
tP(tiC)=false;
ok=1;
while ok
% This while loop tries to add bits to make the
% shortest path wind across as much of the cluster as possible,
% while maintaining the starting and end points
ok=0;
K=size(tiC,2);
idx1=find(tP(tiC(1:2:K-1)+E(2,1:2:K-1))&tP(tiC(2:2:K)+E(2,1:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)-1),tiC(2*idx1(n2)-1)+E(2,2*idx1(n2)-1), tiC(2*idx1(n2))+E(2,2*idx1(n2)-1), tiC(2*idx1(n2):end)];
E=[E(:,1:2*idx1(n2)-2),E([2,1],2*idx1(n2)-1), E([1,2],2*idx1(n2)-1), E([2,1],2*idx1(n2)-1), E(:,2*idx1(n2):end)];
tP(tiC)=false;
end
K=size(tiC,2);
idx1=find(tP(tiC(1:2:K-1)-E(2,1:2:K-1))&tP(tiC(2:2:K)-E(2,1:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)-1),tiC(2*idx1(n2)-1)-E(2,2*idx1(n2)-1), tiC(2*idx1(n2))-E(2,2*idx1(n2)-1), tiC(2*idx1(n2):end)];
E=[E(:,1:2*idx1(n2)-2),E([2,1],2*idx1(n2)-1), E([1,2],2*idx1(n2)-1), E([2,1],2*idx1(n2)-1), E(:,2*idx1(n2):end)];
tP(tiC)=false;
end
K=size(tiC,2);
idx1=find(tP(tiC(2:2:K-1)+E(2,2:2:K-1))&tP(tiC(3:2:K)+E(2,2:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)),tiC(2*idx1(n2))+E(2,2*idx1(n2)), tiC(2*idx1(n2)+1)+E(2,2*idx1(n2)), tiC(2*idx1(n2)+1:end)];
E=[E(:,1:2*idx1(n2)-1),E([2,1],2*idx1(n2)), E([1,2],2*idx1(n2)), E([2,1],2*idx1(n2)), E(:,2*idx1(n2)+1:end)];
tP(tiC)=false;
end
K=size(tiC,2);
idx1=find(tP(tiC(2:2:K-1)-E(2,2:2:K-1))&tP(tiC(3:2:K)-E(2,2:2:K-1)));
for n2=numel(idx1):-1:1
ok=1;
tiC=[tiC(1:2*idx1(n2)),tiC(2*idx1(n2))-E(2,2*idx1(n2)), tiC(2*idx1(n2)+1)-E(2,2*idx1(n2)), tiC(2*idx1(n2)+1:end)];
E=[E(:,1:2*idx1(n2)-1),E([2,1],2*idx1(n2)), E([1,2],2*idx1(n2)), E([2,1],2*idx1(n2)), E(:,2*idx1(n2)+1:end)];
tP(tiC)=false;
end
end
iC=[iC,tiC];
end
end
end
%% neighb_soc
function B = neighb_soc(A,V,nA)
% A is the label matrix BoardCluster, V is board, and nA is the number of
% clusters
[m,n] = size(A);
B=sparse(A(:,1:n-1),A(:,2:n),double(V(:,1:n-1)<V(:,2:n)),nA,nA)...
+sparse(A(:,2:n),A(:,1:n-1),double(V(:,2:n)<V(:,1:n-1)),nA,nA)...
+sparse(A(1:m-1,:),A(2:m,:),double(V(1:m-1,:)<V(2:m,:)),nA,nA)...
+sparse(A(2:m,:),A(1:m-1,:),double(V(2:m,:)<V(1:m-1,:)),nA,nA);
B(1,:) = 0;
B(:,1) = 0;
B=B>0;
end
|
