[ < ] [ > ]   [Contents] [Index] [ ? ]

4.2 Adjoint model and optimisation with Miniker

In the following a possible use of Miniker for optimisation is discussed. More precisely the use of adjoint and control laws in Miniker are presented. Optimisation isn’t the only application of these tools, but it is the most common one. In that case the adjoint may be used to determine the gradient of a functional to perturbations in the control laws, and an optimisation process can use this information to search for the optimum. Another application of the adjoint is to compute the sensitivity of a cost function to parameters (the ones declared in the free_parameters:’ list. Note that the cost function can be sensitive to probe’s variables, even if these are uncoupled with standard variables in the forward calculations; this is the case when minimizing a quadratic distance function between probes (from the model) and the corresponding measurements.

The code is close transcription of the mathematical calculus described in
http://www.lmd.jussieu.fr/ZOOM/doc/Adjoint.pdf . It essentialy reverse time and transpose the four Jacobian matrices: states and transfers are saved in array dimensionned with maxstep Fortran parameter.


[ < ] [ > ]   [Contents] [Index] [ ? ]

4.2.1 Overview of optimisation with Miniker

In the proposed method, Miniker is run twice, one time forward and then backward to determine the trajectory and the adjoint model. After that the control laws are modified by a program external to Miniker. The same steps are repeated until convergence. More pecisely,

forward

The command law h(t) is given (by an explicit law or taken from a file). The trajectory is computed in a classical way, with the additionnal computation of the functional to be optimised, J, prescribed with specific f_set macros. The states, transfers and control laws are stored.

backward

The adjoint variable is computed from the last time T backward. The time increment is re-read as it could have changed during the forward simulation. The system is solved by using the same technics as in the forward simulation, but with a negative time step.

external phase

Now the command should be corrected. This step isn’t covered here, but, for example, minuit the optimisation tool from the CERN could be used. In order to ease such a use of Miniker, the principal program has to be compiled as a subroutine to be driven by an external program (see Calling the model code).

The functionnal J to be optimised is defined as

                       ∫ T
J = ψ [η(T),φ(T),h(T)]+    l[η(τ),φ (τ ),h(τ)]dτ
                        0

Where ψ is the final cost function, l is the integrand cost function and h represents the control laws variations.

The general use of the adjoint model of a system is the determination of the gradient of this J functional to be optimised, with respect to perturbations of the original conditions of the reference trajectory, that is, along its GTLS(6).


[ < ] [ > ]   [Contents] [Index] [ ? ]

4.2.2 Control laws

Each control law is associated with one cell or transfer equation, meaning that a command associated with an equation does not appear in any other equation. It is still possible to add commands acting anywhere by defining a transfer equal to that command.

The control laws associated with states are in the ux_com(.) array, control laws associated with transfers are in the uy_com(.) array. The control laws may be prescribed even when there is no adjoint computed, nor any optimisation, and they are used during simulation, in which case they will act as external sources. To enable the use of commands, the logical flag Zcommand should be .true..

The command can be given either as:

  1. a table of numerical values in the files ‘uxcom.data’ and ‘uycom.data’.
  2. a function of the problem variables. To turn that feature on the logical flag Zlaw should be set to .true. in ‘zinit’. The sequence ‘zcmd_law’ should hold the code filling the ux_com(.) and uy_com(.) arrays, as the code from that sequence is used whenever the control laws are needed. In that case the files ‘uxcom.data’ and ‘uycom.data’ will be filled by the command values generated by the function along the trajectory.

For example in the Lotka-Volterra model, the parameter apar could be a control variable. In that case, apar would be defined as the variable ux_com(1), and either entered as a law in the sequence ‘zcmd_law’ , either written in the file ‘uxcom.data’ step by step. In that case, there must be a perfect corresponodence between time of the commands and time of the run.


[ < ] [ > ]   [Contents] [Index] [ ? ]

4.2.3 Cost function coding and adjoint modeling

First of all the flag zback should be set to .true. in order to allow adjoint model computation:

 
Zback=.true.;

The two functions cout_Psi corresponding with the final cost and cout_l corresponding with the integrand cost are set up with the f_set macros.

Macro: f_set cout_Psi = f(eta(.),ff(.),ux_com(.),uy_com(.))

This macro defines the final cost function. f is a fortran expression which may be function of cell state variables, ‘eta(1)’…‘eta(np)’, transfers ‘ff(1)’…‘ff(mp)’, state control laws ‘ux_com(1)’…‘ux_com(np)’, and transfer control laws ‘uy_com(1)’…‘uy_com(mp)’.

Macro: f_set cout_l = f(eta(.),ff(.),ux_com(.),uy_com(.))

This macro defines the integrand cost function. f is a fortran expression which may be function of cell state variables, ‘eta(1)’…‘eta(np)’, transfers ‘ff(1)’…‘ff(mp)’, state control laws ‘ux_com(1)’…‘ux_com(np)’, and transfer control laws ‘uy_com(1)’…‘uy_com(mp)’.

For example, the following code sets a cost function for the masselottes model:

 
! Initialisation 
  F_set cout_Psi = eta_move(inode:1);
!and f_set cout_l integrand in the functionnal    
  F_set cout_l = 0.;

In that example the functional is reduced to the final value of the first state component. Here, the adjoint vector will correspond to the final sensitivity (at t = 0) of that component (here the first masselotte position) to a perturbation in all initial conditions(7).

The following variables are set during the backward phase, and output in the associated files:

varfileexplanation
v_adj(.)vadj.dataadjoint to eta(.)
w_adj(.)wadj.dataadjoint to ff(.)
wadj(mp+.)gradmuj.dataadjoint to ff(mp+.)
graduej(.)gradxj.dataadjoint to ux_com(.)
gradufj(.)gradyj.dataadjoint to uy_com(.)
hamiltonhamilton.datatime increment, hamiltonian, cost function increment

[ < ] [ > ]   [Contents] [Index] [ ? ]

4.2.4 Sensitivity of cost function to parameters

The sensitivity of the cost function to all the parameters given as arguments of Free_parameters is computed. For the predator model the sensitivity of a cost function consisting in the integral of the predator population with respect with apar an cpar is obtained with a number of parameters set to 2 in ‘dimetaphi’:

 
parameter (lp=2);

And the cost function and Free_parameters list in ‘zinit’:

 
f_set cout_Psi = eta(2);
f_set cout_l = eta(2);
Free_parameters: apar,cpar;

apar and cpar also have to be given a value. The result is output in ‘gradpj.data’.


[Contents] [Index] [ ? ]

This document was generated by a tester on a sunny day using texi2html.