Studies need many runs, not one. These tools reuse, vary and queue experiments for you. All live in Experiments.
Groups: an experiment can belong to a named group of related runs. Copy group, matrix and schedule work on groups.
Experiments, Copy experiment
- Use it for
- Replications, or a baseline you then tweak.
- You provide
- Source, new name, number of copies, group (optional).
- You get
- Copies
name_1,name_2… with configuration, populations, clients and prompts. Fresh database and port: no data, no logs.
Each copy starts from zero.
Experiments, Copy group
- Use it for
- Rerunning a full design, for instance with another model.
- You provide
- Source group and new group name.
- You get
- Every experiment copied as in Copy experiment. If one copy fails you get a report, the others go through.
One click, the whole set.
Experiments, Upload experiment
- Use it for
- A colleague's or archived experiment, or moving between installations.
- You provide
- A ZIP of the experiment folder. Name and group optional.
- You get
- A registered experiment ready to run. Settings are read from
config_server.jsonorserver_config.json, also in nested folders.
No need to re-enter platform or settings.
Experiments, Experiment matrix
- Use it for
- Sensitivity analyses and factorial designs, e.g. recommenders × content filters.
- You provide
- Source experiment, target group, the settings to unlock and alternative values for each.
- You get
- A new group, everything else inherited. Ports, paths and databases are regenerated.
Same inputs, two ways to combine them.
Independent
One experiment per alternative, one setting at a time. For sensitivity analysis.
Cartesian
One experiment per combination. For factorial designs; the count grows fast.
Experiments, Experiment schedule
- Use it for
- Overnight batches, running a whole group.
- You provide
- Groups in the order you want. Dynamic filling is optional.
- You get
- Groups run in sequence, experiments in a group in parallel. Start, stop, progress and logs.
Group 2 starts when group 1 is done.
Good to know
HPC groups run at most 4 experiments at once and are never mixed with Standard. Dynamic filling starts the next run as soon as a slot frees.
Eligible experiments
Stopped, not completed, not already in a group, with no infinite client.
Typical workflow: one experiment, then Matrix into a group, Schedule to run it, Copy group to replicate the study.