Mercurial > hg > camir-aes2014

annotate toolboxes/FullBNT-1.0.7/bnt/examples/dynamic/HHMM/Map/mk_map_hhmm.m @ 0:e9a9cd732c1e tip

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 function bnet = mk_map_hhmm(varargin)
wolffd@0 2
wolffd@0 3 % p is the prob of a successful move (defines the reliability of motors)
wolffd@0 4 p = 1;
wolffd@0 5 obs_model = 'unique';
wolffd@0 6
wolffd@0 7 for i=1:2:length(varargin)
wolffd@0 8 switch varargin{i},
wolffd@0 9 case 'p', p = varargin{i+1};
wolffd@0 10 case 'obs_model', obs_model = varargin{i+1};
wolffd@0 11 end
wolffd@0 12 end
wolffd@0 13
wolffd@0 14
wolffd@0 15 q = 1-p;
wolffd@0 16 unique_obs = strcmp(obs_model, 'unique');
wolffd@0 17
wolffd@0 18 % assign numbers to the nodes in topological order
wolffd@0 19 U = 1; A = 2; C = 3; F = 4;
wolffd@0 20 if unique_obs
wolffd@0 21 onodes = 5;
wolffd@0 22 else
wolffd@0 23 N = 5; E = 6; S = 7; W = 8; % north, east, south, west
wolffd@0 24 onodes = [N E S W];
wolffd@0 25 end
wolffd@0 26
wolffd@0 27 % create graph structure
wolffd@0 28
wolffd@0 29 ss = 4 + length(onodes); % slice size
wolffd@0 30 intra = zeros(ss,ss);
wolffd@0 31 intra(U,F)=1;
wolffd@0 32 intra(A,[C F onodes])=1;
wolffd@0 33 intra(C,[F onodes])=1;
wolffd@0 34
wolffd@0 35 inter = zeros(ss,ss);
wolffd@0 36 inter(U,[A C])=1;
wolffd@0 37 inter(A,[A C])=1;
wolffd@0 38 inter(F,[A C])=1;
wolffd@0 39 inter(C,C)=1;
wolffd@0 40
wolffd@0 41 % node sizes
wolffd@0 42 ns = zeros(1,ss);
wolffd@0 43 ns(U) = 2; % left/right
wolffd@0 44 ns(A) = 2;
wolffd@0 45 ns(C) = 3;
wolffd@0 46 ns(F) = 2;
wolffd@0 47 if unique_obs
wolffd@0 48 ns(onodes) = 5; % we will assign each state a unique symbol
wolffd@0 49 else
wolffd@0 50 ns(onodes) = 2;
wolffd@0 51 end
wolffd@0 52 l = 1; r = 2; % left/right
wolffd@0 53 L = 1; R = 2;
wolffd@0 54
wolffd@0 55 % Make the DBN
wolffd@0 56 bnet = mk_dbn(intra, inter, ns, 'observed', onodes);
wolffd@0 57 eclass = bnet.equiv_class;
wolffd@0 58
wolffd@0 59
wolffd@0 60
wolffd@0 61 % Define CPDs for slice 1
wolffd@0 62 % We clamp all the CPDs that are not tied,
wolffd@0 63 % since we cannot learn them from a single sequence.
wolffd@0 64
wolffd@0 65 % uniform probs over actions (the input could be chosen from a policy)
wolffd@0 66 bnet.CPD{eclass(U,1)} = tabular_CPD(bnet, U, 'CPT', mk_stochastic(ones(ns(U),1)), ...
wolffd@0 67 'adjustable', 0);
wolffd@0 68
wolffd@0 69 % uniform probs over starting abstract state
wolffd@0 70 bnet.CPD{eclass(A,1)} = tabular_CPD(bnet, A, 'CPT', mk_stochastic(ones(ns(A),1)), ...
wolffd@0 71 'adjustable', 0);
wolffd@0 72
wolffd@0 73 % Uniform probs over starting concrete state, modulo the fact
wolffd@0 74 % that corridor 2 is only of length 2.
wolffd@0 75 CPT = zeros(ns(A), ns(C)); % CPT(i,j) = P(C starts in j | A=i)
wolffd@0 76 CPT(1, :) = [1/3 1/3 1/3];
wolffd@0 77 CPT(2, :) = [1/2 1/2 0];
wolffd@0 78 bnet.CPD{eclass(C,1)} = tabular_CPD(bnet, C, 'CPT', CPT, 'adjustable', 0);
wolffd@0 79
wolffd@0 80 % Termination probs
wolffd@0 81 CPT = zeros(ns(U), ns(A), ns(C), ns(F));
wolffd@0 82 CPT(r,1,1,:) = [1 0];
wolffd@0 83 CPT(r,1,2,:) = [1 0];
wolffd@0 84 CPT(r,1,3,:) = [q p];
wolffd@0 85 CPT(r,2,1,:) = [1 0];
wolffd@0 86 CPT(r,2,2,:) = [q p];
wolffd@0 87 CPT(l,1,1,:) = [q p];
wolffd@0 88 CPT(l,1,2,:) = [1 0];
wolffd@0 89 CPT(l,1,3,:) = [1 0];
wolffd@0 90 CPT(l,2,1,:) = [q p];
wolffd@0 91 CPT(l,2,2,:) = [1 0];
wolffd@0 92
wolffd@0 93 bnet.CPD{eclass(F,1)} = tabular_CPD(bnet, F, 'CPT', CPT);
wolffd@0 94
wolffd@0 95
wolffd@0 96 % Observation model
wolffd@0 97 if unique_obs
wolffd@0 98 CPT = zeros(ns(A), ns(C), 5);
wolffd@0 99 CPT(1,1,1)=1; % Theo state 4
wolffd@0 100 CPT(1,2,2)=1; % Theo state 5
wolffd@0 101 CPT(1,3,3)=1; % Theo state 6
wolffd@0 102 CPT(2,1,4)=1; % Theo state 9
wolffd@0 103 CPT(2,2,5)=1; % Theo state 10
wolffd@0 104 %CPT(2,3,:) undefined
wolffd@0 105 O = onodes(1);
wolffd@0 106 bnet.CPD{eclass(O,1)} = tabular_CPD(bnet, O, 'CPT', CPT);
wolffd@0 107 else
wolffd@0 108 % north/east/south/west can see wall (1) or opening (2)
wolffd@0 109 CPT = zeros(ns(A), ns(C), 2);
wolffd@0 110 CPT(:,:,1) = q;
wolffd@0 111 CPT(:,:,2) = p;
wolffd@0 112 bnet.CPD{eclass(W,1)} = tabular_CPD(bnet, W, 'CPT', CPT);
wolffd@0 113 bnet.CPD{eclass(E,1)} = tabular_CPD(bnet, E, 'CPT', CPT);
wolffd@0 114 CPT = zeros(ns(A), ns(C), 2);
wolffd@0 115 CPT(:,:,1) = p;
wolffd@0 116 CPT(:,:,2) = q;
wolffd@0 117 bnet.CPD{eclass(S,1)} = tabular_CPD(bnet, S, 'CPT', CPT);
wolffd@0 118 bnet.CPD{eclass(N,1)} = tabular_CPD(bnet, N, 'CPT', CPT);
wolffd@0 119 end
wolffd@0 120
wolffd@0 121 % Define the CPDs for slice 2
wolffd@0 122
wolffd@0 123 % Abstract
wolffd@0 124
wolffd@0 125 % Since the top level never resets, the starting distribution is irrelevant:
wolffd@0 126 % A2 will be determined by sampling from transmat(A1,:).
wolffd@0 127 % But the code requires we specify it anyway; we make it all 0s, a dummy value.
wolffd@0 128 startprob = zeros(ns(U), ns(A));
wolffd@0 129
wolffd@0 130 transmat = zeros(ns(U), ns(A), ns(A));
wolffd@0 131 transmat(R,1,:) = [q p];
wolffd@0 132 transmat(R,2,:) = [0 1];
wolffd@0 133 transmat(L,1,:) = [1 0];
wolffd@0 134 transmat(L,2,:) = [p q];
wolffd@0 135
wolffd@0 136 % Qps are the parents we condition the parameters on, in this case just
wolffd@0 137 % the past action.
wolffd@0 138 bnet.CPD{eclass(A,2)} = hhmm2Q_CPD(bnet, A+ss, 'Fbelow', F, ...
wolffd@0 139 'startprob', startprob, 'transprob', transmat);
wolffd@0 140
wolffd@0 141
wolffd@0 142
wolffd@0 143 % Concrete
wolffd@0 144
wolffd@0 145 transmat = zeros(ns(C), ns(U), ns(A), ns(C));
wolffd@0 146 transmat(1,r,1,:) = [q p 0.0];
wolffd@0 147 transmat(2,r,1,:) = [0.0 q p];
wolffd@0 148 transmat(3,r,1,:) = [0.0 0.0 1.0];
wolffd@0 149 transmat(1,r,2,:) = [q p 0.0];
wolffd@0 150 transmat(2,r,2,:) = [0.0 1.0 0.0];
wolffd@0 151 %
wolffd@0 152 transmat(1,l,1,:) = [1.0 0.0 0.0];
wolffd@0 153 transmat(2,l,1,:) = [p q 0.0];
wolffd@0 154 transmat(3,l,1,:) = [0.0 p q];
wolffd@0 155 transmat(1,l,2,:) = [1.0 0.0 0.0];
wolffd@0 156 transmat(2,l,2,:) = [p q 0.0];
wolffd@0 157
wolffd@0 158 % Add a new dimension for A(t-1), by copying old vals,
wolffd@0 159 % so the matrix is the same size as startprob
wolffd@0 160
wolffd@0 161
wolffd@0 162 transmat = reshape(transmat, [ns(C) ns(U) ns(A) 1 ns(C)]);
wolffd@0 163 transmat = repmat(transmat, [1 1 1 ns(A) 1]);
wolffd@0 164
wolffd@0 165 % startprob(C(t-1), U(t-1), A(t-1), A(t), C(t))
wolffd@0 166 startprob = zeros(ns(C), ns(U), ns(A), ns(A), ns(C));
wolffd@0 167 startprob(1,L,1,1,:) = [1.0 0.0 0.0];
wolffd@0 168 startprob(3,R,1,2,:) = [1.0 0.0 0.0];
wolffd@0 169 startprob(3,R,1,1,:) = [0.0 0.0 1.0];
wolffd@0 170 %
wolffd@0 171 startprob(1,L,2,1,:) = [0.0 0.0 010];
wolffd@0 172 startprob(2,L,2,1,:) = [1.0 0.0 0.0];
wolffd@0 173 startprob(2,R,2,2,:) = [0.0 1.0 0.0];
wolffd@0 174
wolffd@0 175 % want transmat(U,A,C,At,Ct), ie. in topo order
wolffd@0 176 transmat = permute(transmat, [2 3 1 4 5]);
wolffd@0 177 startprob = permute(startprob, [2 3 1 4 5]);
wolffd@0 178 bnet.CPD{eclass(C,2)} = hhmm2Q_CPD(bnet, C+ss, 'Fself', F, ...
wolffd@0 179 'startprob', startprob, 'transprob', transmat);
wolffd@0 180
wolffd@0 181