Online optimization algorithms 171
beam lifetime, etc. Guided by the response of the machine performance to the
knob changes, the human tuner makes further knob adjustments in order to
maximize the machine performance.
Manual tuning is essentially an optimization process. The function to be
optimized is the machine performance evaluated on the operating machine
through measurements. The knobs are the input variables of the function.
The human tuner executes an optimization algorithm to search the parameter space for the optimum of the performance function. Manual tuning has
many limitations. It is typically slow for humans to dial in the new setpoints,
to process the measured data, and to make decisions on the next move. The
complexity of the optimization problem is usually limited by the ability of
humans to analyze and comprehend the data taken from a high dimension
parameter space in real time. Typically only one knob is tuned at a time.
Efficient search directions that involve multiple knobs cannot be taken advantage of. It is not practical to tune problems with a large number of knobs.
Small trends over a long parameter range could be overlooked, which may
prevent the convergence toward the optimum in some cases. The successful
execution of the manual tuning depends on the tuner’s experience, training,
and familiarity with the system. This could pose a challenge to the training
and retaining of a competent operation team.
In the age of computerized accelerator control, it is clearly desirable to
automate the tuning process. Automated tuning integrates all the three components – knob adjustments, performance monitoring, and decision making –
in one computer program. This not only speeds up the data taking and the
data processing, but also opens up the possibility of using efficient optimization algorithms. Optimization of large scale problems with complex parameter
space (e.g., with strong coupling between the parameters) becomes feasible.
Simultaneous optimization of multiple performance measures is also possible.
Automated tuning has been attempted at many places [35, 32, 34, 2]. The
algorithms employed in these studies include one-dimensional scans, random
optimization, Nelder-Mead simplex, etc. Despite these efforts, automated tuning has not gained much popularity until recently. This was likely due to the
challenges of online optimization that were not met by the traditional optimization algorithms. The power of automated tuning will be manifested only
when algorithms suitable for online optimization are adopted. There is an
ongoing campaign to develop new algorithms for the online tuning of accelerators. As more research is carried out in the area, more effective and efficient
algorithms will appear, which may change the landscape in this emerging field.
7.1.2 Formulating online optimization problem
Online optimization is similar to ordinary mathematical optimization in that
it looks for the maximum or minimum of the objective function(s) within a
certain parameter space. We consider the minimization problem only as any
maximization problem can be turned into minimization by flipping the sign
beam lifetime, etc. Guided by the response of the machine performance to the
knob changes, the human tuner makes further knob adjustments in order to
maximize the machine performance.
Manual tuning is essentially an optimization process. The function to be
optimized is the machine performance evaluated on the operating machine
through measurements. The knobs are the input variables of the function.
The human tuner executes an optimization algorithm to search the parameter space for the optimum of the performance function. Manual tuning has
many limitations. It is typically slow for humans to dial in the new setpoints,
to process the measured data, and to make decisions on the next move. The
complexity of the optimization problem is usually limited by the ability of
humans to analyze and comprehend the data taken from a high dimension
parameter space in real time. Typically only one knob is tuned at a time.
Efficient search directions that involve multiple knobs cannot be taken advantage of. It is not practical to tune problems with a large number of knobs.
Small trends over a long parameter range could be overlooked, which may
prevent the convergence toward the optimum in some cases. The successful
execution of the manual tuning depends on the tuner’s experience, training,
and familiarity with the system. This could pose a challenge to the training
and retaining of a competent operation team.
In the age of computerized accelerator control, it is clearly desirable to
automate the tuning process. Automated tuning integrates all the three components – knob adjustments, performance monitoring, and decision making –
in one computer program. This not only speeds up the data taking and the
data processing, but also opens up the possibility of using efficient optimization algorithms. Optimization of large scale problems with complex parameter
space (e.g., with strong coupling between the parameters) becomes feasible.
Simultaneous optimization of multiple performance measures is also possible.
Automated tuning has been attempted at many places [35, 32, 34, 2]. The
algorithms employed in these studies include one-dimensional scans, random
optimization, Nelder-Mead simplex, etc. Despite these efforts, automated tuning has not gained much popularity until recently. This was likely due to the
challenges of online optimization that were not met by the traditional optimization algorithms. The power of automated tuning will be manifested only
when algorithms suitable for online optimization are adopted. There is an
ongoing campaign to develop new algorithms for the online tuning of accelerators. As more research is carried out in the area, more effective and efficient
algorithms will appear, which may change the landscape in this emerging field.
7.1.2 Formulating online optimization problem
Online optimization is similar to ordinary mathematical optimization in that
it looks for the maximum or minimum of the objective function(s) within a
certain parameter space. We consider the minimization problem only as any
maximization problem can be turned into minimization by flipping the sign