reimplemented gradient ascent as central finite differences method

This commit is contained in:
2026-01-11 12:42:48 -08:00
parent c47b7229ba
commit ec202d7790
12 changed files with 101 additions and 176 deletions

View File

@@ -1,10 +1,11 @@
function obj = run(obj, domain, partitioning, t, index)
function obj = run(obj, domain, partitioning, timestepIndex, index, agents)
arguments (Input)
obj (1, 1) {mustBeA(obj, 'agent')};
domain (1, 1) {mustBeGeometry};
partitioning (:, :) double;
t (1, 1) double;
timestepIndex (1, 1) double;
index (1, 1) double;
agents (:, 1) {mustBeA(agents, 'cell')};
end
arguments (Output)
obj (1, 1) {mustBeA(obj, 'agent')};
@@ -14,134 +15,63 @@ function obj = run(obj, domain, partitioning, t, index)
partitionMask = partitioning == index;
objectiveValues = domain.objective.values(partitionMask); % f(omega) on W_n
% Compute sensor performance across partition
% Compute sensor performance on partition
maskedX = domain.objective.X(partitionMask);
maskedY = domain.objective.Y(partitionMask);
zFactor = 1;
sensorValues = obj.sensorModel.sensorPerformance(obj.pos, obj.pan, obj.tilt, [maskedX, maskedY, zeros(size(maskedX))]); % S_n(omega, P_n) on W_n
sensorValuesLower = obj.sensorModel.sensorPerformance(obj.pos - [0, 0, zFactor * domain.objective.discretizationStep], obj.pan, obj.tilt, [maskedX, maskedY, zeros(size(maskedX))]); % S_n(omega, P_n - [0, 0, z]) on W_n
sensorValuesHigher = obj.sensorModel.sensorPerformance(obj.pos + [0, 0, zFactor * domain.objective.discretizationStep], obj.pan, obj.tilt, [maskedX, maskedY, zeros(size(maskedX))]); % S_n(omega, P_n - [0, 0, z]) on W_n
% Put the values back into the form of the partition to enable basic operations on this data
F = NaN(size(partitionMask));
F(partitionMask) = objectiveValues;
S = NaN(size(partitionMask));
Slower = S;
Shigher = S;
S(partitionMask) = sensorValues;
Slower(partitionMask) = sensorValuesLower;
Shigher(partitionMask) = sensorValuesHigher;
% Find agent's performance
C = S .* F;
obj.performance = [obj.performance, sum(C(~isnan(C)))]; % at current Z only
C = cat(3, Shigher, S, Slower) .* F;
% Compute gradient on agent's performance
[gradCX, gradCY, gradCZ] = gradient(C, domain.objective.discretizationStep); % grad C
gradC = cat(4, gradCX, gradCY, gradCZ);
nGradC = vecnorm(gradC, 2, 4);
if obj.debug
% Compute additional component-level values for diagnosing issues
[gradSensorPerformanceX, gradSensorPerformanceY] = gradient(S, domain.objective.discretizationStep); % grad S_n
[gradObjectiveX, gradObjectiveY] = gradient(F, domain.objective.discretizationStep); % grad f
gradS = cat(3, gradSensorPerformanceX, gradSensorPerformanceY, zeros(size(gradSensorPerformanceX))); % grad S_n
gradF = cat(3, gradObjectiveX, gradObjectiveY, zeros(size(gradObjectiveX))); % grad f
ii = 8;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), F./max(F, [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), S./max(S, [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), vecnorm(gradF, 2, 3)./max(vecnorm(gradF, 2, 3), [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), vecnorm(gradS, 2, 3)./max(vecnorm(gradS, 2, 3), [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), S .* vecnorm(gradF, 2, 3)./max(vecnorm(gradF, 2, 3), [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), F .* vecnorm(gradS, 2, 3)./max(vecnorm(gradS, 2, 3), [], 'all')./(max(F .* vecnorm(gradS, 2, 3)./max(vecnorm(gradS, 2, 3), [], 'all'))));
hold(obj.debugFig.Children(1).Children(ii), "off");
% Compute agent performance at the current position and each delta position +/- X, Y, Z
delta = domain.objective.discretizationStep; % smallest possible step size that gets different results
deltaApplicator = [0, 0, 0; 1, 0, 0; -1, 0, 0; 0, 1, 0; 0, -1, 0; 0, 0, 1; 0, 0, -1]; % none, +X, -X, +Y, -Y, +Z, -Z
C_delta = NaN(7, 1); % agent performance at delta steps in each direction
for ii = 1:7
% Apply delta to position
pos = obj.pos + delta * deltaApplicator(ii, 1:3);
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), C./max(C, [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
ii = ii - 1;
hold(obj.debugFig.Children(1).Children(ii), "on");
cla(obj.debugFig.Children(1).Children(ii));
imagesc(obj.debugFig.Children(1).Children(ii), nGradC./max(nGradC, [], 'all'));
hold(obj.debugFig.Children(1).Children(ii), "off");
[x, y] = find(nGradC == max(nGradC, [], "all"));
% Compute performance values on partition
if ii < 5
% Compute sensing performance
sensorValues = obj.sensorModel.sensorPerformance(obj.pos, obj.pan, obj.tilt, [maskedX, maskedY, zeros(size(maskedX))]); % S_n(omega, P_n) on W_n
% Objective performance does not change for 0, +/- X, Y steps.
% Those values are computed once before the loop and are only
% recomputed when +/- Z steps are applied
else
% Redo partitioning for Z stepping only
partitioning = obj.partition(agents, domain.objective);
% just pick one
r = randi([1, size(x, 1)]);
x = x(r); y = y(r);
% Recompute partiton-derived performance values for objective
partitionMask = partitioning == index;
objectiveValues = domain.objective.values(partitionMask); % f(omega) on W_n
% switch them
temp = x;
x = y;
y = temp;
% find objective location in discrete domain
[~, xIdx] = find(domain.objective.groundPos(1) == domain.objective.X);
xIdx = unique(xIdx);
[yIdx, ~] = find(domain.objective.groundPos(2) == domain.objective.Y);
yIdx = unique(yIdx);
for ii = 8:-1:1
hold(obj.debugFig.Children(1).Children(ii), "on");
% plot GA selection
scatter(obj.debugFig.Children(1).Children(ii), x, y, 'go');
scatter(obj.debugFig.Children(1).Children(ii), x, y, 'g+');
% plot objective center
scatter(obj.debugFig.Children(1).Children(ii), xIdx, yIdx, 'ro');
scatter(obj.debugFig.Children(1).Children(ii), xIdx, yIdx, 'r+');
hold(obj.debugFig.Children(1).Children(ii), "off");
% Recompute partiton-derived performance values for sensing
maskedX = domain.objective.X(partitionMask);
maskedY = domain.objective.Y(partitionMask);
sensorValues = obj.sensorModel.sensorPerformance(pos, obj.pan, obj.tilt, [maskedX, maskedY, zeros(size(maskedX))]); % S_n(omega, P_n) on W_n
end
% Rearrange data into image arrays
F = NaN(size(partitionMask));
F(partitionMask) = objectiveValues;
S = NaN(size(partitionMask));
S(partitionMask) = sensorValues;
% Compute agent performance
C = S .* F;
C_delta(ii) = sum(C(~isnan(C)));
end
% return now if there is no data to work with, and do not move
if all(isnan(nGradC), 'all')
return;
end
% Compute gradient by finite central differences
gradC = [(C_delta(2)-C_delta(3))/(2*delta), (C_delta(4)-C_delta(5))/(2*delta), (C_delta(6)-C_delta(7))/(2*delta)];
% Use largest grad(C) value to find the direction of the next position
[xNextIdx, yNextIdx, zNextIdx] = ind2sub(size(nGradC), find(nGradC == max(nGradC, [], 'all')));
% switch them
temp = xNextIdx;
xNextIdx = yNextIdx;
yNextIdx = temp;
% Compute scaling factor
targetRate = 0.2 - 0.0008 * timestepIndex; % slow down as you get closer
rateFactor = targetRate / norm(gradC);
roundingScale = 10^-log10(domain.objective.discretizationStep);
zKey = zFactor * [1; 0; -1];
pNext = [floor(roundingScale .* mean(unique(domain.objective.X(:, xNextIdx))))./roundingScale, floor(roundingScale .* mean(unique(domain.objective.Y(yNextIdx, :))))./roundingScale, obj.pos(3) + zKey(zNextIdx)]; % have to do some unfortunate rounding here sometimes
% Determine next position
vDir = (pNext - obj.pos)./norm(pNext - obj.pos, 2);
rate = 0.1 - 0.0004 * t; % slow down as you get closer, coming to a stop by the end
nextPos = obj.pos + vDir * rate;
% Compute unconstrained next position
pNext = obj.pos + rateFactor * gradC;
% Move to next position
obj.lastPos = obj.pos;
obj.pos = nextPos;
obj.pos = pNext;
% Reinitialize collision geometry in the new position
d = obj.pos - obj.collisionGeometry.center;