MOP103 Proceedings of LINAC08, Victoria, BC, Canada

ARTIFICIAL INTELLIGENCE RESEARCH IN PARTICLE

ACCELERATOR CONTROL SYSTEMS FOR BEAM LINE TUNING*
Martin Pieck†, Los Alamos National Laboratory, NM 87544, USA

Abstract PREVIOUS AI ATTEPMPTS TO

Tuning particle accelerators is time consuming and ACCELERATOR CONTROL
expensive, with a number of inherently non-linear
interactions between system components. Conventional An early example of AI technology applied to
control methods have not been successful in this domain, accelerator control can be found in Higo et al. [1]. In this
and the result is constant and expensive monitoring of the work a rather simple objective of maintaining a given
systems by human operators. This is particularly true for system or operation condition has been discussed. AI is
the start-up and conditioning phase after a maintenance relevant here for dealing with the complex problems of
period or an unexpected fault. In turn, this often requires a generalized hysteretic and stochastic effects.
step by step restart of the accelerator. Surprisingly few Weygand [2] reports on a development for a
attempts have been made to apply intelligent accelerator knowledge-based, domain specific expert system at
control techniques to help with beam tuning, fault Brookhaven National Laboratory. The purpose of the
detection, and fault recovery problems. The reason for expert system is to aid in the control of the Heavy Ion
that might be that accelerator facilities are rare and Transfer line (HITL) and in turn to minimize down time
difficult to understand systems that require detailed expert after a change in running conditions or the start of a new
knowledge about the underlying physics as well as run. Due to the complexity of the expert system it was
months if not years of experience to understand the divided up into three domains (control, device beam
relationship between individual components, particularly influence, and device cause and effect segment). Separate
if they are geographically disjoint. This paper will give an from that is the goal-solving mechanism of the program.
overview about the research effort in the accelerator This goal-solving function takes a specified high level
community that has been dedicated to the use of artificial goal, and then, by developing a tree of sub-goals, attempts
intelligence methods for accelerator beam line tuning. to solve the given goal via a hill-climbing technique.
However, the system doesn’t take into account that
BEAM LINE TUNING conditions may change during the execution of the
A typical accelerator beam line includes trim magnets solution.
for steering, quadrupole magnets for focusing, Faraday Other attempts at intelligent control for accelerators
cups and stripline detectors for measuring current, and include the ISIS tune advisor (Schultz et al. 1990), the
profile monitors for measuring beam size and position. LAMPF Beam Loss Expert (Clearwater et al. 1986), and a
Beam loss monitors give information about loss of beam learning system based on RL4 (Clearwater et al. 1990).
through miss-steering. Unfortunately, real systems rarely The ISIS tune advisor and LAMPF Beam Loss Expert
work as they are designed. Problems arise from imperfect were both expert systems for indirect control which were
beam production, residual magnetic fields, poorly never implemented as general or real-time control
modeled beam behavior, misplaced or flawed control solutions. The learning system used knowledge-based
elements, and changes to the design or use of the facility induction for off-line learning of beam position monitor
after it has been built. Beam line designers consider these placement, but was not implemented as a general learning
problems and build diagnostic components into the beam algorithm.
lines. Profile monitors and current detectors are used to
measure beam parameters throughout the line to provide Neural networks have been applied to accelerator
information for verifying or correcting beam control for actual manipulation of control parameters as
characteristics. Even so, imperfect detectors, system well as for simulation. Howell et al. [3] used neural
errors, and noise due to various effect cause beam line networks for modeling and control of a negative-ion
control to be difficult at best. Surprisingly, few attempts accelerator source at Los Alamos National Laboratory to
have been made to apply intelligent accelerator control predict the beam characteristics of the source for given
techniques to help with beam tuning, fault detection, and changes in control settings. However, the success was
fault recovery problems. rather limited.
Brown [4] developed an automated controller based on
an artificial neural network and evaluated its applicability
in a real-time environment. This capability was developed
___________________________________________

*This work has benefited from the use of the LANSCE at LANL. This
within the context of a small angle negative ion source on
facility is funded by the US DOE and operated by LANS for NSSA the Discharge Test Stand at Los Alamos National
under Contract DE-AC52-06NA25396. Laboratory. Using no knowledge of operating conditions,
†
[email protected] the controller begins acquiring rough snapshot of the
Extreme Beams and Other Technologies 4D - Beam Dynamics, Computer Simulation, Beam Transport
314
 Proceedings of LINAC08, Victoria, BC, Canada MOP103

operational condition. Once the rough approximation of provided the architecture for integrating multiple pre-
the condition has been determined, the network formulates existing expert systems to exchange information and,
a model from which it determines the best operating point. therefore, increase the overall performance.
The controller takes the ion source to that operating point
for a reality check. As real data is fed in, the model of the Another application of AI techniques in operating a
operating surface is updated until the neural network's larger system is in the field of fault detection and system
model agrees with reality. The controller then uses a diagnosis. Leger et al. (1996) points out that in order to
gradient ascent method to optimize the operation of the operate a successful plant, continues improvement must
ion source. Mead (1994) presents some successful control be made in the areas of safety quality and reliability.
examples of the above application. Central to this continuous improvement is the early or
proactive detection and correct diagnosis of process
The SETUP program developed at CERN (Bouche faults. The paper examines the feasibility of using
1995) is representative of efforts to apply AI techniques to cumulative summation control charts and artificial neural
small subsystems. SETUP is only used for pre-control networks together for fault detection and diagnosis
equipment setup. The program uses an object-oriented (FDD). The control chart and neural network are linked
description language for representing control actions. The by using a characteristic fault signature pattern for each
reasoning system searches the oriented graph defined by fault, which is to be detected and diagnosed.
an object description to make decisions about equipment
setup without human assistance. The program provides a In one of the first papers from Klein and Luger [2] on
good example of using object models for control decision beam tuning, a framework of a general purpose intelligent
making. It does not attempt to perform real-time control control system for a particle accelerator has been
or use on-line feedback from the system. presented. Results of coupling neural network and expert
systems technology to solve several standard accelerator
Arruat [5] reports on symbolic computing programs tuning problems based on realistic simulations have been
that address not only the manipulation of mathematical discussed. The authors also examine the effectiveness of
functions, but also process control. He uses symbolic additional heuristic search techniques such as genetic
computing as a tool for developing algorithmic engines algorithms. However, the algorithm did not converge
that can be fully integrated in the controls environment towards specific solution methods to create the desired
and facilitates modular design of control systems. He result. To solve the problem, the authors also examined
reports that this technique has been applied to a special the capabilities of an analytic control algorithm, fuzzy
class of accelerator problems, namely the beam steering in logic (FL) control, and genetic algorithm searches. It was
transfer lines and accelerator rings in the CERN PS noticed that the accuracy of the FL solution depended
Complex. greatly on the quality of the knowledge the author placed
in the system. The genetic algorithm is an appropriate
In his Ph.D. dissertation Jennings [6] developed and heuristic for focusing control because it can search large
implemented a new model of multi-agent coordination, solution spaces in non-linear domains.
called Joint Responsibility, based upon the notion of joint
intentions. The Responsibility framework was devised In continuation of their earlier work Klein and Luger
specifically for coordinating behavior in complex, (Klein et al. 1997(a)) have identified two different sources
unpredictable and dynamic environments such as particle of control information that must both be incorporated into
accelerator control. Jennings contribution was that he any successful automated control system. The first source
noticed that in a distributed multi-agent system when includes analytic domain knowledge necessary for
anything unexpected happened (i.e. new information modeling the accelerator and beam line. The second,
invalidated existing goals, synchronization between equally important source is experiential knowledge about
actions was disrupted or agents had misinterpreted the the specific facility and group of components being
situation) the multi-agent community would act controlled.
incoherently. This incoherence occurred because agents
did not embody sufficient knowledge about the process of Stern (Stern et al. 1997) reports on the continuing work
team problem solving. Jennings provided agents with an of Klein on coordinating the activity of the control
explicit model of joint problem solving about which they subsystems to adaptively execute sustained sequences of
could reason when deciding how to interact with others. control actions of the type required to tune the whole
This new development found its application in ARCHON. accelerator in a coordinated fashion. The discussed
control system employs a multi-layer organization in
ARCHON (1993-1996) was Europe's largest ever which knowledge–based decision making is used to
project in the area of Distributed Artificial Intelligence dynamically configure a lower level optimization and
(DAI) and applied to CERN control systems for control algorithm. An object-oriented physical access
controlling and diagnosing faults in one of their particle layer supported by a distributed control system allows
accelerators (called the Proton Synchrotron (PS)). The abstraction from the lower level details of hardware
Archon-project, which Jennings was in charge of, manipulation, signal processing, and synchronization.
Extreme Beams and Other Technologies 4D - Beam Dynamics, Computer Simulation, Beam Transport
315
MOP103 Proceedings of LINAC08, Victoria, BC, Canada

This hybrid architecture integrates a variety of down time vs. beam time and try to minimize
methodologies, including teleo-reactive trees for dynamic conditioning, start-up time and fault recovery time. In
exception handling and re-planning. addition, managers increase their attention to preventive
maintenance (predicting when something might happen
In Klein’s last paper [7] he discusses the results of their and then to something beforehand).
research on the use of Teleo-Reactive programming as a
high level control mechanism. According to the authors, Speeding up beam tuning, advanced fault detection, and
control of particle accelerators requires a hybrid preventive maintenance (sometimes referred to as
architecture which includes methodologies for planning, structural health) are areas where AI can and has already
intelligent search, and pattern recognition. The author’s made an impact. Specifically applied to accelerator
suggestion is that control must be distributed and controls, expert beam tuning systems have been around
hierarchical to utilize parallel problem-solving in the face for more than a decade, fault recovery programs got more
of time-sensitive control requirements and to decompose attention at the end of the century. On the other hand
complex control problems into more manageable preventive maintenance has never been applied to
subtasks. As a high level control mechanism, the authors accelerator facilities. So, there is a chance here to lead the
used a planning technique known as Teleo-Reactive (TR) way.
programming developed by Nils Nilsson at Stanford. TR
programming combines the responsiveness of an analog REFERENCES
feedback loop and the goal-oriented behavior of a
[1] Higo T., Shaoee H., and Spencer J.E., “Some
production system, to allow the control system to govern
application of AI to Problems of Accelerator
the execution of its tasks in a dynamic environment. The
Physics”, Proceedings of the IEEE Particle
system has been successfully tested at both the
Accelerator Conference, Washington D.C. (1987).
Brookhaven National Laboratory ATF and at the ATLAS
[2] Weygand D.P., “Artificial Intelligence and
facility at Argonne National Laboratory.
Accelerator Control”, Proceedings of the IEEE
Particle Accelerator Conference, Washington DC,
Fiesler [8] reports on a suit of tests on predictive
(1987).
algorithms to determine their ability to learn from and
[3] Howell, J. A., et al.., “Control of a negative-ion
find relationships between large numbers of variables. In
accelerator source using neural networks “, Nuclear
particular, they used COMFORT, a particle accelerator
Instruments and Methods in Physics Research
simulator, to generate large amounts of data. They
Section A, Volume 293, Issue 1-2, p. 517-522.
eventually compared the results among several
[4] Brown, S. K., et al., “Optimization and control of a
fundamentally different types of algorithms, including
small angle ion source using an adaptive neural
least squares and hybrid neural networks with real data
network controller (invited) “, Review of Scientific
that were obtained from Brookhaven National Laboratory.
Instruments, Volume 65, Issue 4, April 1994,
pp.1411-1415.
SUMMARY [5] Arruat M., et al., “Integration of symbolic computing
It seems surprising that virtually all research on AI in accelerator control”, International Journal of
applied to particle accelerator stopped around the year Modern Physics C (Physics and Computers); Aug.
2000. This certainly can’t be explained with the Y2K 1995; vol.6, no.4, p.475-80, Fourth International
problem. Nevertheless, it seems like interest and/or Workshop on Software Engineering, Artificial
funding must have stopped around that time. The Intelligence, and Expert Systems for High Energy
alternative is, of course, that the research in the area and Nuclear Physics, 3-8 April 1995, Pisa, Italy
wasn’t promising or the problem was so complex that [6] Jennings N. R., et al., "Using ARCHON to develop
researchers did not find the return of investment very real-word DAI applications for electricity
appealing to them. transportation management and particle accelerator
control", IEEE Expert, (1996).
Nevertheless, the author of this paper sees a great [7] Klein, W.B., et al., "Teleo-Reactive Control for
chance for AI to make its comeback in accelerator control Accelerator Beamline Tuning". In Proceedings of the
rooms. The first and foremost reason is that running these IASTED International Conference. Zurich:
machines has become a time and cost critical issue. The IASTED/ACTA Press (2000).
older an accelerator gets the more maintenance/upgrade [8] Fiesler, E., Campbell, S., “Hybrid neural networks
work needs to be invested. This is in turn creates a and their application to particle accelerator control”,
problem since users/experimenters are paying to do their Proceedings of the SPIE - The International Society
experiments and extensive maintenance/upgrade work for Optical Engineering; 1999; vol.3812, p.132-42,
reduces experimental time. With less time for experiments Proceedings of the 1999 Applications and Sciences of
less money comes in for maintenance. However, less Neural Networks, Fuzzy Systems, and Evolutionary
maintenance means that the accelerators become more Computation II; Jul 19-Jul 20 1999; Denver, CO,
error/fault prone. Therefore, managers have to balance USA

Extreme Beams and Other Technologies 4D - Beam Dynamics, Computer Simulation, Beam Transport
316