Difference between revisions of "AIOP Jun. 3, 2024"

From epsciwiki
Jump to navigation Jump to search
 
(One intermediate revision by the same user not shown)
Line 67: Line 67:
 
=== Minutes ===
 
=== Minutes ===
 
<!-- Attendees: Torri J., Thomas B., David L., Patrick M., Armen K. -->
 
<!-- Attendees: Torri J., Thomas B., David L., Patrick M., Armen K. -->
 +
[https://docs.google.com/document/d/17JP7JY1d6IFqUYTYyHMKfak6nzSwtQG_99hJH9p7Kmk/edit?usp=sharing Google doc]
 +
  <h2>Project Progress</h2>
 +
   
 +
    <h3>Patrick</h3>
 +
    <ul>
 +
        <li>Taking a look at bad CPP runs, specifically 101617-101618, there are a host of logbook entries noting problems:
 +
            <ul>
 +
                <li>[https://logbooks.jlab.org/entry/4030558 Shift Summary]</li>
 +
                <li>[https://logbooks.jlab.org/entry/4030574 Coherent Edge Pic]</li>
 +
                <li>[https://logbooks.jlab.org/entry/4030723 Description of Rates Being Off]</li>
 +
            </ul>
 +
        </li>
 +
        <li>We may need to go back and refit the coherent peak with good initial parameters.</li>
 +
        <li>Work with Richard Jones:
 +
            <ul>
 +
                <li>Might have a good way to reconstruct the spectra from the archived data using TAGM/TAGH rates from Aluminum and then can do the refitting.</li>
 +
                <li>David suggested using larger statistics if we wanted.</li>
 +
            </ul>
 +
        </li>
 +
        <li>Answering Thomas’ question about shift taker behavior, namely how long between button presses:
 +
            <ul>
 +
                <li>Patrick showed a histogram of the time between button presses, most of which is skewed towards lower values between 0-30 seconds. In the figure, a nudge is anytime the set value of the pitch or yaw changes in a one-second time step.</li>
 +
                <li>The scalers are read back from the discriminators every second, while the fit result takes longer and also probably has a sleep function built into it.</li>
 +
                <li>If the shift crew starts nudging and then the beam goes off and then they continue to nudge when the beam comes back, it is treated as the same sequence.</li>
 +
                <li>Naomi suggests stopping the clock when the beam goes out, and Patrick requires the beam to be up for at least 30 seconds to keep the event.</li>
 +
            </ul>
 +
        </li>
 +
    </ul>
 +
   
 +
    <h3>Jiawei</h3>
 +
    <ul>
 +
        <li>No slides this week.</li>
 +
        <li>Converging data sets, currently the discrepancy has been reduced to an excess of 7% of nudge events in Jiawei’s dataset compared to Patrick's.</li>
 +
        <li>Working with Hovanes and Richard Jones on a method to locate the beam spot on the diamond.
 +
            <ul>
 +
                <li>Hovanes thinks it would be good if there is some description of how this is supposed to work. It all depends on which part of the diamond is being irradiated too much, and the beam widths are not tracked aside from log entries from the harp scans.</li>
 +
                <li>For the software to work, the widths should be recorded in some reasonable place. Harp scan data can be piped into text files such that we can analyze those.</li>
 +
                <li>Nominal coherent peak positions for Fall 2018 data is not recorded in wiki pages, so parsing RC notes and logbooks. David will send Jiawei links to see the white board pictures.</li>
 +
            </ul>
 +
        </li>
 +
    </ul>
 +
   
 +
    <h3>Armen</h3>
 +
    <ul>
 +
        <li>Meeting on Friday to go over RL efforts into source side of the project – data science group has toolkits for RL already.</li>
 +
        <li>If the angle can be calculated from the shift in energy, we don’t need AI for that.</li>
 +
    </ul>
 +
   
 +
    <h3>Patrick</h3>
 +
    <ul>
 +
        <li>Cristiano and I have also started some offline RL algorithm (using SCOPE-RL) for small set of data.</li>
 +
    </ul>
 +
   
 +
    <h3>Malachi</h3>
 +
    <ul>
 +
        <li>Environment needs to be established, and OpenAI’s gymnasium is industry standard that we should adhere to. Happy to help with teaching existing frameworks in use across multiple applications.</li>
 +
        <li>Hovanes asks questions about training RL models:
 +
            <ul>
 +
                <li>RL learns from experience. With historical data, we can train a surrogate model to mimic interactions learned from data (so we don’t forget things about the past) and then use that to interact and train the policy.</li>
 +
            </ul>
 +
        </li>
 +
    </ul>
  
  
 
==== Action Items ====
 
==== Action Items ====
 +
 +
<ul>
 +
        <li>Find harp scan data files and get them into a usable format (Naomi suggests deferring).</li>
 +
        <li>Establish data set we are going to give to ML and ensure the data is in a suitable form.</li>
 +
<li> Establish who will be presenting at PTSP2024 at JLab [https://indico.jlab.org/event/823/ Indico]
 +
        <li>ML-focused meeting on Friday, using the regular target zoom link!</li>
 +
    </ul>
 +
 +
=== Milestones! ===
 +
<ul>
 +
<li>Identify relevant parameters </li>
 +
<li>Identify nudge events </li>
 +
<li><strong>Shapley analysis!</strong> </li>
 +
</ul>

Latest revision as of 19:44, 3 June 2024

The current weekly meeting time is every other Monday at 13:00 US/Eastern

Connection Info:

You can connect using ZOOM Video conferencing (ID: 161 588 8669). (Click "Expand" to the right for details -->):

One tap mobile: US: +16692545252,,1602484178# or +16468287666,,1602484178#

Meeting URL: https://jlab-org.zoomgov.com/j/1615888669?pwd=MnhGVUhpeEd3NVpnTGxoZHJManNBdz09&from=addon
Meeting ID: 161 588 8669
Passcode: 561441

Join by Telephone
For higher quality, dial a number based on your current location.
Dial: +1 669 254 5252 US (San Jose)
+1 646 828 7666 US (New York)
+1 646 964 1167 US (US Spanish Line)
+1 551 285 1373 US (New Jersey)
+1 669 216 1590 US (San Jose)
+1 415 449 4000 US (US Spanish Line)
833 568 8864 US Toll-free
Meeting ID: 161 588 8669
International numbers
Join from an H.323/SIP room system
H.323: 161.199.138.10 (US West)
161.199.136.10 (US East)
Meeting ID: 161 588 8669
Passcode: 561441

SIP: 1615888669@sip.zoomgov.com
Passcode: 561441

One tap mobile: US: +16692545252,,1601987443# or +16468287666,,1601987443#

Meeting URL: https://jlab-org.zoomgov.com/j/1615888669?pwd=MnhGVUhpeEd3NVpnTGxoZHJManNBdz09&from=addon
Meeting ID: 161 588 8669
Passcode: 561441

AIOP - Polarized Photon Source

Agenda:

  1. Previous Meeting
  2. Announcements
  3. Project Progress
  4. PIER Activities
    • Middle School Data Science Hackathon Workshop
    • HUGS mini-Workshop
  5. AOT


Minutes

Google doc

Project Progress

Patrick

  • Taking a look at bad CPP runs, specifically 101617-101618, there are a host of logbook entries noting problems:
  • We may need to go back and refit the coherent peak with good initial parameters.
  • Work with Richard Jones:
    • Might have a good way to reconstruct the spectra from the archived data using TAGM/TAGH rates from Aluminum and then can do the refitting.
    • David suggested using larger statistics if we wanted.
  • Answering Thomas’ question about shift taker behavior, namely how long between button presses:
    • Patrick showed a histogram of the time between button presses, most of which is skewed towards lower values between 0-30 seconds. In the figure, a nudge is anytime the set value of the pitch or yaw changes in a one-second time step.
    • The scalers are read back from the discriminators every second, while the fit result takes longer and also probably has a sleep function built into it.
    • If the shift crew starts nudging and then the beam goes off and then they continue to nudge when the beam comes back, it is treated as the same sequence.
    • Naomi suggests stopping the clock when the beam goes out, and Patrick requires the beam to be up for at least 30 seconds to keep the event.

Jiawei

  • No slides this week.
  • Converging data sets, currently the discrepancy has been reduced to an excess of 7% of nudge events in Jiawei’s dataset compared to Patrick's.
  • Working with Hovanes and Richard Jones on a method to locate the beam spot on the diamond.
    • Hovanes thinks it would be good if there is some description of how this is supposed to work. It all depends on which part of the diamond is being irradiated too much, and the beam widths are not tracked aside from log entries from the harp scans.
    • For the software to work, the widths should be recorded in some reasonable place. Harp scan data can be piped into text files such that we can analyze those.
    • Nominal coherent peak positions for Fall 2018 data is not recorded in wiki pages, so parsing RC notes and logbooks. David will send Jiawei links to see the white board pictures.

Armen

  • Meeting on Friday to go over RL efforts into source side of the project – data science group has toolkits for RL already.
  • If the angle can be calculated from the shift in energy, we don’t need AI for that.

Patrick

  • Cristiano and I have also started some offline RL algorithm (using SCOPE-RL) for small set of data.

Malachi

  • Environment needs to be established, and OpenAI’s gymnasium is industry standard that we should adhere to. Happy to help with teaching existing frameworks in use across multiple applications.
  • Hovanes asks questions about training RL models:
    • RL learns from experience. With historical data, we can train a surrogate model to mimic interactions learned from data (so we don’t forget things about the past) and then use that to interact and train the policy.


Action Items

  • Find harp scan data files and get them into a usable format (Naomi suggests deferring).
  • Establish data set we are going to give to ML and ensure the data is in a suitable form.
  • Establish who will be presenting at PTSP2024 at JLab Indico
  • ML-focused meeting on Friday, using the regular target zoom link!

Milestones!

  • Identify relevant parameters
  • Identify nudge events
  • Shapley analysis!
Retrieved from "https://wiki.jlab.org/epsciwiki/index.php?title=AIOP_Jun._3,_2024&oldid=5073"