170 Beam-based Correction and Optimization for Accelerators
dial in the trial solutions and to perform analysis on the diagnostic data and
employing advanced online optimization algorithms, automated tuning can
deal with bigger and more complex problems. It not only can expedite the
routine tuning processes, but more importantly, can solve accelerator tuning
problems that cannot be accomplished with other methods.
Online optimization algorithms are essential to the strength of automated
tuning. Because of the measurement errors that enter the performance metrics and other characteristics of the online application environment, online
optimization algorithms need to address special challenges. Traditional optimization algorithms that are powerful for smooth functions might not be
suitable for online problems. New algorithms that were modified from traditional algorithms with simple-minded improvements may be powerful tools.
In this chapter we will first discuss the general considerations of the online optimization approach. This is followed by a review of the optimization
algorithms for continuous functions. Analytic functions with added random
errors are used to test the performance of the algorithms for online applications. The robust conjugate direction search (RCDS) method is introduced
and compared to the other algorithms through the tests.
7.1 GENERAL CONSIDERATIONS OF ONLINE OPTIMIZATION
7.1.1 Need for online optimization of accelerators
Beam-based correction is a satisfying approach. It seems natural to accelerator
physicists as it employs the physics principles that govern the machines. The
correction process is deterministic. The success of the correction approach
signifies that accelerator physicists understand and have control over the inner
workings of the machine.
However, the correction approach is not always applicable. For some machines or some applications, there are no diagnostics available to detect the
discrepancies between the machine and the ideal target. In some cases, the
diagnostics cannot provide the detailed information needed to solve for the
required parameter adjustments. In other cases, the diagnostics simply do
not exist. There may be a lack of a target for correction. If there are major differences between the model and the real machine, the target suggested
by the model may not be applicable. In this case, the target itself needs to
be discovered. It is also possible that both diagnostics and the target exist,
but there is no reliable, deterministic method to deduce the solution for the
correction toward the target. In these circumstances, empirically tuning the
control parameters for the optimal performance is a sensible choice.
Manual tuning is frequently employed in accelerator control rooms by accelerator physicists and operators. During manual tuning, one or a few control
parameters (i.e., knobs) are adjusted while the machine performance is monitored. Depending on the applications, the machine performance indicator may
be beam intensity, transmission or capture efficiency, beam sizes, beam loss,
dial in the trial solutions and to perform analysis on the diagnostic data and
employing advanced online optimization algorithms, automated tuning can
deal with bigger and more complex problems. It not only can expedite the
routine tuning processes, but more importantly, can solve accelerator tuning
problems that cannot be accomplished with other methods.
Online optimization algorithms are essential to the strength of automated
tuning. Because of the measurement errors that enter the performance metrics and other characteristics of the online application environment, online
optimization algorithms need to address special challenges. Traditional optimization algorithms that are powerful for smooth functions might not be
suitable for online problems. New algorithms that were modified from traditional algorithms with simple-minded improvements may be powerful tools.
In this chapter we will first discuss the general considerations of the online optimization approach. This is followed by a review of the optimization
algorithms for continuous functions. Analytic functions with added random
errors are used to test the performance of the algorithms for online applications. The robust conjugate direction search (RCDS) method is introduced
and compared to the other algorithms through the tests.
7.1 GENERAL CONSIDERATIONS OF ONLINE OPTIMIZATION
7.1.1 Need for online optimization of accelerators
Beam-based correction is a satisfying approach. It seems natural to accelerator
physicists as it employs the physics principles that govern the machines. The
correction process is deterministic. The success of the correction approach
signifies that accelerator physicists understand and have control over the inner
workings of the machine.
However, the correction approach is not always applicable. For some machines or some applications, there are no diagnostics available to detect the
discrepancies between the machine and the ideal target. In some cases, the
diagnostics cannot provide the detailed information needed to solve for the
required parameter adjustments. In other cases, the diagnostics simply do
not exist. There may be a lack of a target for correction. If there are major differences between the model and the real machine, the target suggested
by the model may not be applicable. In this case, the target itself needs to
be discovered. It is also possible that both diagnostics and the target exist,
but there is no reliable, deterministic method to deduce the solution for the
correction toward the target. In these circumstances, empirically tuning the
control parameters for the optimal performance is a sensible choice.
Manual tuning is frequently employed in accelerator control rooms by accelerator physicists and operators. During manual tuning, one or a few control
parameters (i.e., knobs) are adjusted while the machine performance is monitored. Depending on the applications, the machine performance indicator may
be beam intensity, transmission or capture efficiency, beam sizes, beam loss,