I am rewriting a Monte Carlo simulation model in MATLAB with an emphasis on readability. The model involves many particles, represented as (x,y,z), following a random walk over a small set of states with certain termination probabilities. The information relevant for output is the number of particles that terminate in a given state.
The simulation requires enough particles that running it for each particle individually is cost prohibitive. Vectorization seems to be the way to get performance out of MATLAB, but is there any idiomatic way of creating a vectorized version of this simulation in MATLAB?
I'm beating my head against the wall to accomplish this - I've even tried creating a (nStates x nParticles) matrix representing each particle-state combination, but this approach quickly spirals out of control in terms of readability since particles bounce from state to state independently of one another. Should I just bite the bullet and switch to a language more suitable for this?
Just write the code as you normally would. Almost all matlab functions can accept and return vectorized input. For instance, to simulate a brownian motion of N particles in 1 dimension
position = zeros([N 1]); %start at origin
sigma = sqrt(D * dt); %D is diffusion coefficient, dt is time step
for j = 1:numSteps
position = position + sigma*randn(size(position));
end
if you wanted to have a different sigma for each position, you would make sigma a vector the same size as position and use "dot times" notation to indicate element by element operation
position = position + sigma.*randn(size(position));
if the scattering was an arbitrary function of position and some random element, you would just have to write a vectorized function, e.g.
function newstep = step(position)
%diffusion in a overdamped harmonic potential
newstep = -dt*k*position + D*randn(size(position));
for j = 1:numsteps; position = position + step(position);
and so on
Related
I am working on a MR-physic simulation written in Matlab which simulates bloch's equations on an defined object. The magnetisation in the object is updated every time-step with the following functions.
function Mt = evolveMtrans(gamma, delta_B, G, T2, Mt0, delta_t)
% this function calculates precession and relaxation of the
% transversal component, Mt, of M
delta_phi = gamma*(delta_B + G)*delta_t;
Mt = Mt0 .* exp(-delta_t*1./T2 - 1i*delta_phi);
end
This function is a very small part of the entire code but is called upon up to 250.000 times and thus slows down the code and the performance of the entire simulation. I have thought about how I can speed up the calculation but haven't come up with a good solution. There is one line that is VERY time consuming and stands for approximately 50% - 60% of the overall simulation time. This is the line,
Mt = Mt0 .* exp(-delta_t*1./T2 - 1i*delta_phi);
where
Mt0 = 512x512 matrix
delta_t = a scalar
T2 = 512x512 matrix
delta_phi = 512x512 matrix
I would be very grateful for any suggestion to speed up this calculation.
More info below,
The function evovleMtrans is called every timestep during the simulation.
The parameters that are used for calling the function are,
gamma = a constant. (gyramagnetic constant)
delta_B = the magnetic field value
G = gradientstrength
T2 = a 512x512 matrix with T2-values for the object
Mstart.r = a 512x512 matrix with the values M.r had the last timestep
delta_t = a scalar with the difference in time since the last calculated M.r
The only parameters of these that changed during the simulation are,
G, Mstart.r and delta_t. The rest do not change their values during the simulation.
The part below is the part in the main code that calls the function.
% update phase and relaxation to calcTime
delta_t = calcTime - Mstart_t;
delta_B = (d-d0)*B0;
G = Sq.Gx*Sq.xGxref + Sq.Gz*Sq.zGzref;
% Precession around B0 (z-axis) and B1 (+-x-axis or +-y-axis)
% is defined clock-wise in a right hand system x, y, z and
% x', y', z (see the Bloch equation, Bloch 1946 and Levitt
% 1997). The x-axis has angle zero and the y-axis has angle 90.
% For flipping/precession around B1 in the xy-plane, z-axis has
% angle zero.
% For testing of precession direction:
% delta_phi = gamma*((ones(size(d)))*1e-6*B0)*delta_t;
M.r = evolveMtrans(gamma, delta_B, G, T2, Mstart.r, delta_t);
M.l = evolveMlong(T1, M0.l, Mstart.l, delta_t);
This is not a surprise.
That "single line" is a matrix equation. It's really 1,024 simultaneous equations.
Per Jannick, that first term means element-wise division, so "delta_t/T[i,j]". Multiplying a matrix by a scalar is O(N^2). Matrix addition is O(N^2). Evaluating exponential of a matrix will be O(N^2).
I'm not sure if I saw a complex argument in there as well. Does that mean complex matricies with real and imaginary entries? Does your equation simplify to real and imaginary parts? That means twice the number of computations.
Your best hope is to exploit symmetry as much as possible. If all your matricies are symmetric, you cut your calculations roughly in half.
Use parallelization if you can.
Algorithm choice can make a big difference, too. If you're using explicit Euler integration, you may have time step limitations due to stability concerns. Is that why you have 250,000 steps? Maybe a larger time step is possible with a more stable integration schema. Think about a higher order adaptive scheme with error correction, like 5th order Runge Kutta.
There are several possibilities to improve the speed of the code but all that I see come with a caveat.
Numerical ode integration
The first possibility would be to change your analytical solution by numerical differential equation solver. This has several advantages
The analytical solution includes the complex exponential function, which is costly to calculate, while the differential equation contains only multiplication and addition. (d/dt u = -a u => u=exp(-at))
There are plenty of built-in solvers for matlab available and they are typically pretty fast (e.g. ode45). The built-ins however all use a variable step size. This improves speed and accuracy but would be a problem if you really need a fixed equally spaced grid of time points. Here are unofficial fixed step solvers.
As a start you could also try to use just an euler step by replacing
M.r = evolveMtrans(gamma, delta_B, G, T2, Mstart.r, delta_t);
by
delta_phi = gamma*(delta_B + G)*t_step;
M.r += M.r .* (1-t_step*1./T2 - 1i*delta_phi);
You can then further improve that by precalculating all constant values, e.g. one_over_T1=1/T1, moving delta_phi out of the loop.
Caveat:
You are bound to a minimum step size or the accuracy suffers. Therefore this is only a good idea if you time-spacing is quite fine.
Less points in time
You should carfully analyze whether you really need so many points in time. It seems somewhat puzzling to me that you need so many points. As you know the full analytical solution you can freely choose how to sample the time and maybe use this to your advantage.
Going fortran
This might seem like a grand step but in my experience basic (simple loops, matrix operations etc.) matlab code can be relatively easily translated to fortran line-by-line. This would be especially helpful in addition to my first point. If you still want to use the full analytical solution probably there is not much to gain here because exp is already pretty fast in matlab.
What is the fitness function used to solve an inverted pendulum ?
I am evolving neural networks with genetic algorithm. And I don't know how to evaluate each individual.
I tried minimize the angle of pendulum and maximize distance traveled at the end of evaluation time (10 s), but this won't work.
inputs for neural network are: cart velocity, cart position, pendulum angular velocity and pendulum angle at time (t). The output is the force applied at time (t+1)
thanks in advance.
I found this paper which lists their objective function as being:
Defined as:
where "Xmax = 1.0, thetaMax = pi/6, _X'max = 1.0, theta'Max =
3.0, N is the number of iteration steps, T = 0.02 * TS and Wk are selected positive weights." (Using specific values for angles, velocities, and positions from the paper, however, you will want to use your own values depending on the boundary conditions of your pendulum).
The paper also states "The first and second terms determine the accumulated sum of
normalised absolute deviations of X1 and X3 from zero and the third term when minimised, maximises the survival time."
That should be more than enough to get started with, but i HIGHLY recommend you read the whole paper. Its a great read and i found it quite educational.
You can make your own fitness function, but i think the idea of using a position, velocity, angle, and the rate of change of the angle the pendulum is a good idea for the fitness function. You can, however, choose to use those variables in very different ways than the way the author of the paper chose to model their function.
It wouldn't hurt to read up on harmonic oscillators either. They take the general form:
mx" + Bx' -kx = Acos(w*t)
(where B, or A may be 0 depending on whether or not the oscillator is damped/undamped or driven/undriven respectively).
I am using virtual potential fields to control the movement of a group of robots in a 2D environment, their position is given by a matrix of x and y coordinates. The virtual potential fields depend on a number of variables, one of them is the inter-robot distance. A short (heavily simplified) example of my code is given below.
x = sym('x',[4 2]); % four robots with x and y coordinates
xd = sym('xd',[1 2]); % a single destination
F = sym(ones(4,1)); % one potential function for each robot
for i=1:size(x,1)
for j=1:size(x,1)
if i~=j
F(i) = F(i)/norm(x(i,:)-x(j,:))^2; % infinite potential when any two robots collide
end
end
F(i) = F(i) * norm(x(i,:)-xd)^2; % add an attraction force to the goal
end
So now that I have created symbolic expressions for the potential fields, I need to find their derivative so I can apply steepest descent. Now I'm wondering: does it make any difference whether I use the function gradient or diff to obtain the derivative with respect to the position? To clarify: for robot i I want to take the derivative with respect to xi_1 and xi_2.
Your question, as stated, is bordering on mathematics rather than programming. The gradient is just the generalization of the derivative to multiple dimensions. Yes, for movement in a 2-D plane, it would make sense to use sym/gradient. As the documentation states, if you specify just a scalar for the second argument, sym/gradient becomes equivalent to sym/diff. To properly calculate your 2-D gradient, the second argument must be a two-element vector, e.g., [xi_1 xi_2].
I have attempted to write a code in order to solve the following coupled partial differential EM wave equations:
The code employs finite difference time domain using the Yee algorithm which can be read about in the following two online documents:
http://www.eecs.wsu.edu/~schneidj/ufdtd/ufdtd.pdf
http://www.eecs.wsu.edu/~schneidj/ufdtd/chap3.pdf
I want my source at the left hand side boundary to be a sinusoidal wave with parameters as such:
Ex(0,t) = E0 sin(2πft) for 0 ≤ t ≤ 1/f
The code updates the electric and magnetic field properties of the wave with each loop.
My initial code is as follows:
%FDTD Yee algorithm to solve coupled EM wave equations
clear
clc
G=50; %Specify size of the grid
f=10^3; %choose initial frequency of wave in hertz
e=1; %specify permitivity and permeability (normalised condition)
u=1;
Nt=150; %time steps
E0=1; %Electric Field initial amplitude
%Specify step sizes using corruant condition
c=3*10^8;
dx=0.01;
dt=dx/2*c;
%make constant terms
c1=-dt/(dx*e);
c2=-dt/(dx*u);
%create vgector place holders
Ex=zeros(1,G);
Hy=zeros(1,G);
%create updating loop
M=moviein(Nt);
for t=1:Nt
% Spatial Ex
for k=2:G-1
Ex(k)=Ex(k)+c1*(Hy(k)-Hy(k-1));
end
Ex(G)=0; %PEC boundary condition
%E Source at LHS boundary
Ex(1)=E0*sin(2*pi*f*t);
%Spatial Hy
for n=1:G-1
Hy(n)=Hy(n)+c2*(Ex(n)-Ex(n+1));
end
Hy(G)=0; %PMC boundary condition
plot(Ex);
M(:,t) = getframe;
end
movie(M,1);
Basically I want to create an updating movie which shows the sinusoidal wave propagating to the right hand side boundary coded as a perfect electrical conductor; therefore reflecting the wave, and then propagating back to the left hand side boundary coded as a perfect insulator; absorbing the wave.
The problems I have are as follows:
1) I'm not sure how to properly implement my desired source. It don't appear to be purely sinusoidal.
2) The wave I've coded begins to propagate but it quickly disappears for the majority of the simulation. I do not know why this is occurring
3) I do not know much about running a movie simulation and the plot oscillates as the solution is being run. How can I stop this?
Your wave attenuates because the diference equations are not correctly implemented; instead:
Ex(k)=Ex(k)+c1*(Hy(k)-Hy(k-1));
you should use
Ex1(k)=Ex(k)+c1*(Hy(k)-Hy(k-1));
and instead of:
Hy(n)=Hy(n)+c2*(Ex(n)-Ex(n+1));
you should use:
Hy1(n)=Hy(n)+c2*(Ex(n)-Ex(n+1));
and, in the end of the loop update the current "dataframe":
Hy = Hy1;
Ey = Ey1;
(you should take care also the boundary conditions).
If you want a fixed plot frame that doesn't change when your data changes, create first a axis where you can plot into, with a defined xmin/max and ymin/max, see http://www.mathworks.com/help/matlab/ref/axis.html
You should set the Courant number closer to 1 say 0.995. Thus delta_t = 0.995*delta_x/c.
Also assuming delta_x is in METRIC units then e and u should be in metric units.
I do not know about the specific coding language used but in c or c++ there is no need for intermediate variable Ey1 etc.
Also there should be at least 10 samples per wavelength for accuracy ( preferably 60). Thus wavelength = 60*delta_x and thus the frequency equals roughly of the order 10 to power of 9. Also, I think the sinesoidal source should be E0 * sin(2* pi * f * t * delta_t). You need to adjust your constants, and try it again
Please can you help me understand how to calculate the Mean-Squared Displacement for a single particle moving randomly within a given period of time. I have read a lot of articles on this (including Saxton,1991,Single-Particle Tracking: The Distribution of Diffusion Coefficients), but still confused (not getting the right answer).
Let me start by showing you how I do it and please correct me if I'm wrong:
The way I'm doing it is as follows:
1.Run the program from t=0 to t=100
2.Calculate the displacement, (s(t)-s(t+tau)), at each timestep (ie. at t=1,2,3,...100) and store it in a vector
3.Square the answer to number 2
4.find the mean to the answer of 3
In essence, this is what I'm doing in Matlab
%Initialise the lattice with a square consisting of 16 nonzero lattice sites then proceed %as follows to calculate the MSD:
for t=1:tend
% Allow the particle to move randomly in the lattice. Then do the following
[row,col]=find(lattice>0);
centroid=mean([row col]);
xvec=[xvec centroid(2)];
yvec=[yvec centroid(1)];
k=length(xvec)-1; % Time
dt=1;
diffx = xvec(1:k) - xvec((1+dt):(k+dt));
diffy = yvec(1:k) - yvec((1+dt):(k+dt));
xsquare = diffx.^2;
ysquare = diffy.^2;
MSD=mean(xsquare+ysquare);
end
I'm trying to find the MSD in order to compute the diffusion co-efficient. Note that I'm modelling a collection of lattice sites (16) to represent a single particle (more biologically realistic), instead of just one. I have been brief with the comment within the for loop as it is quite long, but I'm happy to send it to you.
So far, I'm getting very small MSD values (in the range of 0.001-1), whereas I'm supposed to get values in the range of (10-50). The particle moves very large distances so surely my range of 0.001-1 cannot be right!
This is an extract from the article which I'm trying to reproduce their figure:
" We began by running some simulations in 1D for a single
cell. We allowed the cell to move for a given number of
Monte Carlo time steps (MCS), worked out the mean square
distance traveled in that time, repeated this process 500
times, and evaluate the mean squared distance for this t.
We then repeated this process ten times to get the mean of
. The reason for this choice of repetitions was to
keep the time required to run the simulations within a reasonable
level yet ensuring that the standard deviation of the
mean was relatively small (<7%)".
You can access the article here "From discrete to a continuous model of biological cell movement, 2004, by Turner et al., Physical Review E".
Any hints are greatly appreciated.
How many dimensions does the particle move along ?
I don't have Matlab right now, but here is how I'd do that over one dimension :
% pos is the vector of positions
delta = pos(2:100) - pos(1:99);
meanSquared = mean(delta .* delta);
First of all, why have a particle cover multiple lattice sites? What counts for MSD, in the end, is the displacement of the centroid, which can be represented as a point. If your particle (or cell) is large, or only takes large steps, you can always just make a wider grid. Also, if you're trying to reproduce a figure from somewhere else, you should really use the same algorithm.
For your Monte Carlo simulation, what do you do? If all you really want is get a displacement, you can generate a bunch of random movement vectors in one go (using rand or randi), and use cumsum to calculate the positions. Also, have you plotted your random walks to make sure the data is sensible?
Then, your code looks a bit funny (see comments). Why don't you just use the code provided in this answer to calculate MSD from the positions?
for t=1:tend
% Allow the particle to move randomly in the lattice. Then do the following
[row,col]=find(lattice>0); %# what do you do this for?
centroid=mean([row col]);
xvec=[xvec centroid(2)];
yvec=[yvec centroid(1)]; %# till here, I have no idea what you want to do
k=length(xvec)-1; % Time %# you should subtract dt here
dt=1; %# dt should depend on t!
diffx = xvec(1:k) - xvec((1+dt):(k+dt));
diffy = yvec(1:k) - yvec((1+dt):(k+dt));
xsquare = diffx.^2;
ysquare = diffy.^2;
MSD=mean(xsquare+ysquare);
end