Network Models of Expertise in the Complex Task of Operating Particle Accelerators

Aashwin Ananda Mishra; Jane Shtalenkova; Roussel Rahman; Wan-Lin Hu

REVIEW 4 major objections 5 minor 13 references

Network Models of Expertise in the Complex Task of Operating Particle Accelerators

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Expertise in accelerator tuning is mostly rewiring, not regrouping.

desk verdict A plausible new way to measure expertise from operator logs, but the experience gradient is entangled with calendar time and the NLP filter is unvalidated; still worth a serious referee. read the letter →

arxiv 2412.17988 v1 pith:TSWQOUEW submitted 2024-12-23 cs.SI cs.SYeess.SYstat.AP

Roussel Rahman , Jane Shtalenkova , Aashwin Ananda Mishra , Wan-Lin Hu This is my paper · ORCID

classification cs.SIcs.SYeess.SYstat.AP

keywords humanlearningcomplextasknetworkmodelsexpertiseparticleacceleratorsoperatorlogsnaturallanguageprocessingcommunitydetection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that expertise in an extremely complex real-world task—tuning the X-ray brightness of the Linac Coherent Light Source free-electron laser—can be observed as rewiring in a network of subtasks built from 14 years of operator log entries. The authors claim that as operators gain experience, the importance of individual subtasks, the strength of connections between them, and the whole-network structure all change continuously, while the way subtasks are grouped into communities stays essentially fixed. If true, this would mean learning in such tasks is a divide-and-conquer process: all operators break the task into the same manageable parts, and expertise consists of learning which connections within and across those parts to strengthen. The payoff is a practical, data-driven way to study strategy change in real-world work settings where experiments are impractical.

What carries the argument

The central object is an undirected weighted network whose 27 nodes are the FEL tuning parameters (the subtasks) and whose edge weights are the co-occurrence strengths of parameter mentions within individual electronic-log entries. Entries are first filtered as tuning-related by cosine similarity to an expert-written training article in a 100-topic Latent Semantic Indexing space, then each parameter is identified from the article's parameter sections, and an adjacency matrix is accumulated from within-entry collocations. The argument is carried by comparing networks across experience bins using PageRank for node importance, edge weights for interconnections, the Louvain algorithm with modularity for community structure, hierarchical clustering for within-community organization, and adjacency/Laplacian spectral distances for whole-network change.

What would settle it

Record the accelerator control system's actual parameter-change events during tuning sessions and build the same kind of weighted network from those events; if the edge weights from action data do not correlate with the edge weights from elog co-occurrence, then the co-occurrence proxy fails, and the central rewiring claim is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that operator expertise in FEL tuning manifests as a stable three-community partition of the 27 tuning parameters—beam transport, beam energy/compression, and miscellaneous—paired with systematic changes in how the parameters are used and connected. Novices, intermediates, and experts all produce the same communities, with only one or two parameters classified differently, even though the number of possible partitions is astronomically large. What changes with expertise are the PageRank importance of individual tuning parameters, the edge-weight distribution (experts show larger ratios of out-of-community to in-community weights, 0.43 vs 0.27 for novices), the hierarchical distances at which subtasks converge into communities, and the overall spectral distances between networks. Experts connect subtasks according to accelerator-physics relationships, such as linking BC1 compression to undulator launch, while novices connect geographically close subtasks. The conclusion is that operators of all stages adopt a common divide-and-conquer decomposition of the task, and expertise is a matter of reweighting and rewiring within that stable structure.

Load-bearing premise

The load-bearing assumption is that co-occurrence of two tuning-parameter mentions within a single operator log entry reflects the strength of interconnection between those subtasks in the operator's task strategy or mental model; if co-occurrence instead reflects logging conventions, report templates, or the training article's organization, the claimed expertise-related changes in interconnection structure do not follow.

Editorial extensions

If this is right

Novices initially bind subtasks that lie close together in the machine, whereas experts bind subtasks that are physically distant but causally connected, so training that teaches these long-range physics links could accelerate the transition.
Because the three communities are stable across expertise, improvements in task strategy come from reweighting and rewiring within and across those communities, not from discovering new groupings of subtasks.
The ratio of out-of-community to in-community edge weights grows from 0.27 for novices to 0.43 for experts, suggesting that expertise involves integrating across functional groups rather than staying within them.
Network-level spectral distances rise continuously with years of experience, so the whole-task representation provides a single quantitative signature of operator seniority that could be tracked over time.
The methodology is designed to be generalizable: operational logs are common in real-world human-machine environments, so the same pipeline could map strategy change in other complex tasks without laboratory experiments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the stable three-community partition reflects the training article's organization or the machine's physical layout, then novices may acquire the grouping before they learn optimal within-group sequencing; this could be tested by comparing the detected communities with the section structure of the training materials.
The finding that expertise reweights rather than regroups suggests a concrete, testable training intervention: teach novices the specific long-range cause-effect connections experts use (e.g., BC1 compression to undulator launch) and measure whether the edge-weight distribution and whole-network distance shift toward the expert pattern.
Replacing the co-occurrence proxy with the accelerator control system's recorded parameter-change events would provide a direct falsification check; the paper itself flags action data as future work, and this is the natural next experiment.
If co-occurrence is partly an artifact of logging conventions or report templates, then expertise differences might reflect changing logging style rather than changing strategy; comparing networks from free-form entries versus structured/template entries would separate these.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a real attempt to measure expertise passively from operator logs, and the stable-communities result is interesting enough to chase. But the 'expertise' axis is not cleanly separated from calendar time or logging style, and the current evidence is descriptive rather than demonstrative. I'd send it to review, with the expectation that the authors can address the confound with the data they promise.

What is genuinely new: using 14 years of accelerator elogs to build co-occurrence networks of 27 tuning subtasks, then tracking changes at four levels. The setup is sensible—training article defines the node set, log entries supply the edges, and the community partition is checked with operators. The main empirical finding, that the same three communities persist across expertise levels while edge weights and PageRank shift, is a coherent and nontrivial observation. The authors are also unusually candid in Section 7.1 about the elog being incomplete, logging style being variable, and novices being supervised.

The soft spots are real but mostly addressable. First, the biggest threat is the calendar-era confound. The stress-test overstates the case when it says experienced operators necessarily wrote before 2018—an operator who started in 2017 reaches four years in 2021, so expert-stage entries can be recent. Still, the mix of hire dates and the 14-year span mean that 'years of experience' and 'calendar year' are not independent, and no analysis conditions on entry year. That is a load-bearing issue for the expertise interpretation. Second, the NLP filter is unvalidated: the 0.3 cosine threshold selects ~2,000 of 132,000 entries, with no precision/recall or sensitivity analysis. Third, the paper reports trends without confidence intervals; bootstrapping is mentioned but not shown, so it is hard to tell which changes are signal. The co-occurrence-as-cognition assumption is the deepest assumption, but it is explicitly stated and the authors acknowledge it; given the promise of code and data, a skeptic can re-analyze.

Who is this for: people working on cognitive network science, human-AI teaming, or passive measurement of expertise from operational logs. It is not a definitive test of bounded rationality, but it is a useful proof of concept.

Recommendation: accept for peer review, conditional on the authors adding calendar-year controls, validating the entry-selection step, and reporting uncertainty. I would not desk-reject this.

Referee Report

4 major / 5 minor

Summary. The paper introduces a network-based method to study expertise in the complex real-world task of FEL tuning at the LCLS particle accelerator. It processes 14 years of operator electronic log entries, filters tuning-related entries using LSI topic modeling and a cosine-similarity threshold of 0.3, identifies 27 tuning parameters by reference to a training article, and constructs weighted networks from co-occurrences of parameter mentions within entries. The networks are analyzed at four levels—node (PageRank), edge (edge weights), community (Louvain detection), and whole network (spectral distances)—and compared across years of experience and across three expertise groups (Novice, Intermediate, Expert). The central claim is that subtask communities remain remarkably stable across expertise, while node-level, edge-level, and whole-network structure change consistently, suggesting that operators share a common divide-and-conquer grouping but differ in how they select and interconnect subtasks.

Significance. If the results hold, the paper offers a practical, non-invasive methodology for studying expertise in complex operational tasks using naturally occurring log data. The use of 14 years of operator logs, the multi-level network analysis, and the public release of code and data are notable strengths, as is the domain-expert validation of the detected community labels. The approach could generalize to other complex sociotechnical settings where operational logs are available. However, the current evidence does not yet secure the central claim because of a potentially load-bearing calendar-time confound and an unvalidated NLP pipeline. The paper is best viewed as a promising methodological demonstration whose empirical conclusions require substantially stronger validation.

major comments (4)

[§5.1, §6.1, Figure 4] The expertise variable, defined as time since the operator's first elog entry in Section 5.1, is cross-sectionally confounded with calendar era: because the archive spans 2009–2022, operators with more than 4 years of experience necessarily wrote most of their entries before 2018, while novices wrote mainly in 2021–2022. The monotonic increases in the distance curves in Figure 4 and the group differences in Figure 5 may therefore reflect historical drift in accelerator configuration, tuning procedures, logging software, or documentation norms rather than individual learning. Section 7.1 acknowledges that the elog record is incomplete and highly variable, but the authors provide no analysis controlling for calendar year, entry date, or logging style. I ask the authors to show that the trends in Figure 4 and the group differences in Figure 5 persist when comparison is restricted to overlapping calendar years, or to include entry year as a covariate.
[§5.2–5.3] The NLP pipeline uses a cosine-similarity threshold of 0.3 to select approximately 2,000 tuning-related entries from about 132,000 preprocessed entries, and the same training article is used both to select entries and to identify the 27 parameter mentions. No precision, recall, or human-coded validation is reported for either the entry-selection step or the parameter-identification step. Because all network nodes, edges, and downstream measures depend on this step, the authors should provide validation statistics on a hand-labeled sample and a sensitivity analysis over the similarity threshold and the number of LSI topics (k=100). The absence of such validation makes it difficult to assess whether the constructed networks faithfully represent tuning subtask co-occurrence.
[§5.3, §6.3] Edge weights are defined by co-occurrence of parameter mentions within an elog entry and are interpreted in Section 6.3 as the 'strength of interconnection between subtasks in the mental model of operators.' The paper's own limitations section (7.1) states that the elog record is not a complete description of actions, that individual differences in syntax and event choice make the dataset highly variable, and that novice actions are supervised by experts. Without validation against an independent behavioral measure (e.g., time-stamped control-system actions, actual tuning sequences, or expert judgments of task relatedness), the rewiring claims—such as experts linking BC1 compression to Undulator Launch—remain speculative. At minimum, the authors should soften the mental-model interpretation or provide converging evidence from another data stream.
[§6.1, §7.1] Figure 4 presents the distance curves at all four network levels without confidence intervals, error bars, or significance tests. Section 7.1 mentions that bootstrapping methods were used to check the stability of results, but no bootstrapped intervals, null distributions, or permutation-test results are reported. Given the modest sample (67 operators, 2,088 entries) and the highly unequal group sizes in Table 1 (e.g., 26 experts contributing 822 entries vs. 54 novices contributing 390), the observed differences could be within sampling variability. The central claim of 'consistent changes with expertise' at the subtask, interconnection, and whole-task levels needs to be supported by explicit uncertainty quantification.

minor comments (5)

[Section 4] The sentence beginning 'he operators undergo on-the-job accelerator physics...' contains a typo; 'he' should be 'The'.
[Section 6.4] The text states that edge weights 'remain quite similar across the four groups,' but the main-text comparison involves three expertise groups (Novice, Intermediate, Expert); the fourth group, if it refers to all operators combined, should be identified explicitly.
[Section 6.4] The cross-reference 'In section 6.1, we will demonstrate the changes in the edge weight distributions with each year of experience' should be corrected: Section 6.1 precedes Section 6.4 and reports those results, so the sentence should refer to already-presented results rather than future demonstrations.
[Section 5.4.4] The description of the hierarchical clustering method says the authors adopted 'complete' linkage using the average distance between all pairs of points across two groups; this definition corresponds to average linkage, not complete linkage. The terminology should be corrected to match the described procedure.
[Section 5.3] The phrase 'we use the tuning article – this time broken into the 27 parameters – to identify the topics of entries' is ambiguous about the mechanism of parameter identification; a more precise description of how parameter mentions are extracted (e.g., keyword matching vs. topic attribution) would improve clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: network measures emerge from elog collocation data, and the expertise axis is a time-based grouping rather than a fitted outcome.

full rationale

The paper's derivation chain is self-contained with respect to the circularity patterns checked. Networks are built in Section 5.3 from parameter collocations in operator elog entries; the FEL tuning training article is used only to select tuning-relevant entries and to identify the 27 parameter names, not to prescribe edge weights, community structure, or expertise differences. The expertise variable (Section 5.1) is time elapsed since an operator's first elog entry, not a fitted outcome, and the network measures (PageRank, edge-weight distributions, Louvain communities, spectral distances) are computed from the constructed adjacency matrices rather than derived from the expertise labels. The reported stability of three communities and the expertise-related changes emerge from the collocation data; the consultation with domain experts is a post-hoc interpretation, not an input to the networks. Self-citations to Rahman and Gray support background claims about method changes and are not load-bearing for the network construction or the expertise comparison. The reader's concern about calendar-era confounding is a validity risk rather than circularity, because it does not make any prediction equivalent to its input by construction; Section 7.1 also openly acknowledges the elog record's incompleteness and the expert supervision of novices. No fitted parameter is renamed as a prediction, and no load-bearing claim reduces to a self-citation or to the training article's structure.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or theoretical entities are introduced; the networks are analytical models built from existing text data. The main assumptions are the proxy measures listed above.

free parameters (6)

Cosine similarity threshold for tuning-entry classification = 0.3
Used in Section 5.3 to decide which elog entries are about FEL tuning; directly determines the ~2,000-entry working dataset. No sensitivity analysis is reported.
Minimum entry word count = 10 words
Entry filtering criterion in Section 5.3; arbitrary and not justified empirically.
Number of LSI topics = 100
Dimensionality of the topic space in Section 5.3 and Appendix B; standard but hand-chosen.
Modularity threshold for 'strong' communities = 0.30
Criterion in Sections 5.4.3 and 6.3 taken from Newman (2004); used to declare partitions strong.
Expertise group boundaries = 1 and 4 years
Table 1; bins said to align with operator training stages, but no independent performance validation is provided.
Reference period for distance measures = first six months
Section 6.1; all year-wise changes are measured relative to the first half-year network, so the baseline is itself a novice network.

assumptions (5)

domain assumption The 27 tuning parameters listed in the FEL tuning training article constitute the complete set of relevant subtasks.
Introduced in Section 4 ('Operators have identified twenty-seven different parameters...') and Appendix A; the networks contain only these nodes. If other important tuning actions exist, the network is incomplete.
domain assumption Co-occurrence of parameter mentions in an elog text entry indicates a meaningful relationship between those subtasks in the operator's strategy.
Used in Section 5.3 to build the adjacency matrix; this is the core measurement model and is untested against actual operator actions.
domain assumption Elapsed time since an operator's first elog entry is a valid proxy for expertise.
Defined in Section 5.1 ('we calculated the experience of the operator ... by the time elapsed since the operator wrote their first elog entry'); no performance metric validates this proxy.
domain assumption LSI with cosine similarity can accurately distinguish FEL-tuning-related entries and parameters despite a 3,500-word reference article.
Section 5.3; no precision/recall validation is reported, and the authors note the reference text is short.
standard math Standard graph-theoretic definitions (PageRank, modularity, spectral distance) apply meaningfully to the constructed collocation networks.
Section 5.4; these are standard measures, but their interpretation as cognitive structure is an additional assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Network Models of Expertise in the Complex Task of Operating Particle Accelerators." pith.science (2026). https://pith.science/paper/TSWQOUEW

@misc{pith2026241217988,
  author       = {Pith},
  title        = {Pith review of: Network Models of Expertise in the Complex Task of Operating Particle Accelerators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSWQOUEW}},
  note         = {Machine review of arXiv:2412.17988}
}

read the original abstract

We implement a network-based approach to study expertise in a complex real-world task: operating particle accelerators. Most real-world tasks we learn and perform (e.g., driving cars, operating complex machines, solving mathematical problems) are difficult to learn because they are complex, and the best strategies are difficult to find from many possibilities. However, how we learn such complex tasks remains a partially solved mystery, as we cannot explain how the strategies evolve with practice due to the difficulties of collecting and modeling complex behavioral data. As complex tasks are generally networks of many elementary subtasks, we model task performance as networks or graphs of subtasks and investigate how the networks change with expertise. We develop the networks by processing the text in a large archive of operator logs from 14 years of operations using natural language processing and machine learning. The network changes are examined using a set of measures at four levels of granularity - individual subtasks, interconnections among subtasks, groups of subtasks, and the whole complex task. We find that the operators consistently change with expertise at the subtask, the interconnection, and the whole-task levels, but they show remarkable similarity in how subtasks are grouped. These results indicate that the operators of all stages of expertise adopt a common divide-and-conquer approach by breaking the complex task into parts of manageable complexity, but they differ in the frequency and structure of nested subtasks. Operational logs are common data sources from real-world settings where people collaborate with hardware and software environments to execute complex tasks, and the network models investigated in this study can be expanded to accommodate multi-modal data. Therefore, our network-based approach provides a practical way to investigate expertise in the real world.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

[1]

Aggarwal, P., Gonzalez, C., & Dutt, V. (2016). Cyber-security: Role of deception in cyber-attack detection. Advances in Human Factors in Cybersecurity: Proceedings of the AHFE 2016 International Conference on Human Factors in Cybersecurity, July 27-31, 2016, Walt Disney World®, Florida, USA, 85–96. Anderson, J. R. (1982). Acquisition of cognitive skill.Ps...

work page 2016
[3]

The other notations are: Ni = N (Ci), Nj = N (Cj), N(i,j) = N (Ci∩ Cj), and N = N (Ci∪ Cj)

whereCi and Cj refer to the clusters in partitionsP1 and P2 respectively. The other notations are: Ni = N (Ci), Nj = N (Cj), N(i,j) = N (Ci∩ Cj), and N = N (Ci∪ Cj). ARI(P1, P2) = ∑ (i,j) (N(i,j) 2 ) − (∑ i (Ni 2 ) ×∑ j (Nj 2 ) )/ (N 2 ) 1 2 (∑ i (Ni 2 ) +∑ j (Nj 2 ) )− (∑ i (Ni 2 ) ×∑ j (Nj 2 ) )/ (N 2 ) (C6) Adjusted Mutual Information or AMIbetweentwop...

work page 2009
[31]

electron

EXPERTISE IN PARTICLE ACCELERATORS 31 Appendix A The 27 FEL Tuning Subtasks Table A1 List of FEL tuning parameters ID Parameters 0 LASER iris position 1 Gun Solenoid Strength 2 Schottky Phase 3 CQ01 SQ01 4 XC01 YC01 5 135 MeV bunch length 6 LASER Pulse Stacker Delay 7 LASER heater e- beam overlap (3D) 8 LASER Heater Power 9 Injector Matching Quads QA01 QA...

work page 2019
[58]

A., Salas, E., & Converse, S

Cannon-Bowers, J. A., Salas, E., & Converse, S. (1993). Shared mental models in expert team decision making.Individual and group decision making: Current issues, 221, 221–46. Chase, W. G., & Simon, H. A. (1973). Perception in chess.Cognitive psychology, 4(1), 55–81. Colletier, J.-P., Sawaya, M. R., Gingery, M., Rodriguez, J. A., Cascio, D., Brewster, A. S...

arXiv 1993
[273]

A., Edelen, A., Hanuka, A., & Mayes, C

Mishra, A. A., Edelen, A., Hanuka, A., & Mayes, C. (2021). Uncertainty quantification for deep learning in particle accelerator applications.Physical Review Accelerators and Beams, 24(11), 114601. Nathan, M. J., & Sawyer, R. K. (2014). Foundations of the learning sciences. Newell, A., & Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the law...

work page 2021
[369]

R., Matessa, M., & Lebiere, C

Anderson, J. R., Matessa, M., & Lebiere, C. (1997). Act-r: A theory of higher level cognition and its relation to visual attention.Human–Computer Interaction, 12(4), 439–462. Anderson, J. R., & Schunn, C. D. (2013). Implications of the act-r learning theory: No magic bullets. InAdvances in instructional psychology, volume 5(pp. 1–33). Routledge. Andrews, ...

work page 1997
[376]

Manning, C. D. (2008).Introduction to information retrieval. Cambridge university press. Mathieu, J. E., Heffner, T. S., Goodwin, G. F., Salas, E., & Cannon-Bowers, J. A. (2000). The influence of shared mental models on team process and performance.Journal of applied psychology, 85(2),

work page 2008
[523]

F., & Lebiere, C

Gonzalez, C., Lerch, J. F., & Lebiere, C. (2003). Instance-based learning in dynamic decision making. Cognitive Science, 27(4), 591–635. Gray, W. D. (2017). Plateaus and asymptotes: Spurious and real limits in human performance. Current Directions in Psychological Science, 26(1), 59–67. Gray, W. D., & Lindstedt, J. K. (2017). Plateaus, dips, and leaps: Wh...

work page 2003

Show all 13 references

[1089]

Pellegrini, C., & Stöhr, J. (2003). X-ray free-electron lasers—principles, properties and applica- tions. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 500(1-3), 33–40. Rahman, R. (2022). Dynamic...

2003
[1381]

Gigerenzer, G. (2020b). What is bounded rationality? InRoutledge handbook of bounded ratio- nality (pp. 55–69). Routledge. Gigerenzer, G., & Goldstein, D. G. (1996). Reasoning the fast and frugal way: Models of bounded rationality.Psychological Review, 103(4), 650–669. Goedsch...

2020
[1970]

Simon,H.A.,&Reed,S.K.(1976).Modelingstrategyshiftsinaproblem-solvingtask

American Psychologist, 26(2), 145–159. Simon,H.A.,&Reed,S.K.(1976).Modelingstrategyshiftsinaproblem-solvingtask. Cognitive Psychology, 8(1), 86–97. Simon, H. A. (1997). Models of bounded rationality: Empirically grounded economic reason (Vol. 3). MIT press. Sun, R. (2008). Int...

1976
[2009]

The principles behind the indices are extremely simple to explain

among many other resources due to their popularity. The principles behind the indices are extremely simple to explain. Suppose we have a binary classification problem (Yes/No). For each element labeled positive (i.e., a “Yes”), the label may be correct (True positive, TP) or i...

2016
[2019]

A., & Schaeffer, J

Simon, H. A., & Schaeffer, J. (1992). The game of chess.Handbook of Game Theory With Economic Applications, 1, 1–17. Simon, H. A. (1962). The architecture of complexity.Proceedings of the American Philosophical Society, 106(6), 467–482. Simon, H. A. (1976). From substantive to...

1992

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Aggarwal, P., Gonzalez, C., & Dutt, V. (2016). Cyber-security: Role of deception in cyber-attack detection. Advances in Human Factors in Cybersecurity: Proceedings of the AHFE 2016 International Conference on Human Factors in Cybersecurity, July 27-31, 2016, Walt Disney World®, Florida, USA, 85–96. Anderson, J. R. (1982). Acquisition of cognitive skill.Ps...

work page 2016

[2] [3]

The other notations are: Ni = N (Ci), Nj = N (Cj), N(i,j) = N (Ci∩ Cj), and N = N (Ci∪ Cj)

whereCi and Cj refer to the clusters in partitionsP1 and P2 respectively. The other notations are: Ni = N (Ci), Nj = N (Cj), N(i,j) = N (Ci∩ Cj), and N = N (Ci∪ Cj). ARI(P1, P2) = ∑ (i,j) (N(i,j) 2 ) − (∑ i (Ni 2 ) ×∑ j (Nj 2 ) )/ (N 2 ) 1 2 (∑ i (Ni 2 ) +∑ j (Nj 2 ) )− (∑ i (Ni 2 ) ×∑ j (Nj 2 ) )/ (N 2 ) (C6) Adjusted Mutual Information or AMIbetweentwop...

work page 2009

[3] [31]

electron

EXPERTISE IN PARTICLE ACCELERATORS 31 Appendix A The 27 FEL Tuning Subtasks Table A1 List of FEL tuning parameters ID Parameters 0 LASER iris position 1 Gun Solenoid Strength 2 Schottky Phase 3 CQ01 SQ01 4 XC01 YC01 5 135 MeV bunch length 6 LASER Pulse Stacker Delay 7 LASER heater e- beam overlap (3D) 8 LASER Heater Power 9 Injector Matching Quads QA01 QA...

work page 2019

[4] [58]

A., Salas, E., & Converse, S

Cannon-Bowers, J. A., Salas, E., & Converse, S. (1993). Shared mental models in expert team decision making.Individual and group decision making: Current issues, 221, 221–46. Chase, W. G., & Simon, H. A. (1973). Perception in chess.Cognitive psychology, 4(1), 55–81. Colletier, J.-P., Sawaya, M. R., Gingery, M., Rodriguez, J. A., Cascio, D., Brewster, A. S...

arXiv 1993

[5] [273]

A., Edelen, A., Hanuka, A., & Mayes, C

Mishra, A. A., Edelen, A., Hanuka, A., & Mayes, C. (2021). Uncertainty quantification for deep learning in particle accelerator applications.Physical Review Accelerators and Beams, 24(11), 114601. Nathan, M. J., & Sawyer, R. K. (2014). Foundations of the learning sciences. Newell, A., & Rosenbloom, P. S. (1981). Mechanisms of skill acquisition and the law...

work page 2021

[6] [369]

R., Matessa, M., & Lebiere, C

Anderson, J. R., Matessa, M., & Lebiere, C. (1997). Act-r: A theory of higher level cognition and its relation to visual attention.Human–Computer Interaction, 12(4), 439–462. Anderson, J. R., & Schunn, C. D. (2013). Implications of the act-r learning theory: No magic bullets. InAdvances in instructional psychology, volume 5(pp. 1–33). Routledge. Andrews, ...

work page 1997

[7] [376]

Manning, C. D. (2008).Introduction to information retrieval. Cambridge university press. Mathieu, J. E., Heffner, T. S., Goodwin, G. F., Salas, E., & Cannon-Bowers, J. A. (2000). The influence of shared mental models on team process and performance.Journal of applied psychology, 85(2),

work page 2008

[8] [523]

F., & Lebiere, C

Gonzalez, C., Lerch, J. F., & Lebiere, C. (2003). Instance-based learning in dynamic decision making. Cognitive Science, 27(4), 591–635. Gray, W. D. (2017). Plateaus and asymptotes: Spurious and real limits in human performance. Current Directions in Psychological Science, 26(1), 59–67. Gray, W. D., & Lindstedt, J. K. (2017). Plateaus, dips, and leaps: Wh...

work page 2003