These files together with the NDFA-package will run simulations
presented in the Neural Computation by Valpola & Karhunen (2002).

The NDFA-package is assumed to be found from the directory ndfa.

Start the simulation by run_lorenz.m.  It loads the data from
lorenzdata.mat and initializes the network using init_em.m.  For the
first 500 iterations, the data is five-fold: x(t-2) ... x(t+2) instead
of x(t) at time t.  After that, the extra parts are left out and
the model is saved.

Then another script, more_lorenz.m, is started.  It runs until 1
milj. iterations have been made.  Results are saved every 500
iterations.  At this point the simulation can be stopped and restarted
by more_lorenz.m immediately after stopping the simulation by ctrl-C or
after the last model has been loaded from file.

See the file ndfa_iter.m of the NDFA-package to see how to interpret
the variables in the workspace after running the simulation.

The data file lorenzdata.mat contains the following variables:
  data         10x1000         80000  double array
  genmlp        1x1             7266  struct array
  latents       5x1000         40000  double array
  noi          10x1000         80000  double array
  orla          8x1000         64000  double array
genmlp is the MLP network used for mapping the latents to data, noi is the
noise added to the data.  orla are the original latent variables which are
the states of two indpendent Lorenz processes and one harmonic oscillator.

The file lorenzdata2000.mat contains the variables
  data         10x2000        160000  double array
  latents       5x2000         80000  double array
  noi          10x2000        160000  double array
  orla          8x2000        128000  double array
that is 1000 more datapoint compared to lorenzdata.mat.  They can be
used for testing the prediction performance of the model.

The files nanpca.m, covnan.m and sumnan.m are used by the embedding
initialization.

---

The exact results depend on the platform since in long simulations,
small differences in the last decimals of floating point numbers
gradually build up.  The results depend at least on the Matlab version
and the versions of math libraries of the computer system.
