Training and Predicting with Model Committees¶

This tutorial describes how training and prediction in RuNNer 2 can be extended to incorporate committees (ensembles of neural networks) for improved robustness and uncertainty estimation.

Committee Training¶

RuNNer supports training multiple committee members simultaneously within a single run.

Because all committee members share the feature calculation, they must also share:

• The same dataset
• The same feature definitions
• The same train/test split

RuNNer supports training a committee of models all at once. However, because all committee members share the feature calculation, the train / test split has to be identical for all members.

Most output files generated during training are written into automatically created subdirectories. The folder structure will look like this:

PATH/
├── input.nn
├── scaling.data
├── 1/
│   ├── weights_short.001.data
│   └── weights_short.008.data
├── 2/
│   ├── weights_short.001.data
│   └── weights_short.008.data
├── .../ (other committee folders)
└── n/
    ├── weights_short.001.data
    └── weights_short.008.data

Each directory corresponds to one committee member. Notice how the scaling.data file is written to the root directory of the calculation, since features are shared between committee members.

Prerequisites¶

This tutorial assumes you already have a working single-network training setup.

If not, please start with the 2G tutorial before proceeding.

Input¶

Compared to standard (non-committee) training, only minimal modifications to input.nn are required.

Most importantly we have to specify the number of committee members we want to use with num_committee_members:

num_committee_members 4

This instructs RuNNer to train four neural networks simultaneously using identical architecture and training settings.

Random Seed Behavior¶

If you provide a single random_seed, it is used to * initialize the train/test split * initialize the first committee member * deterministically generate random seeds for all remaining committee members

See the random_seed documentation for further details.

Output¶

Because multiple networks are trained at once, the STDOUT printed by RuNNer expands to show metrics for each member.

--------------------------------------------------------------------------------
                             Energy [meV / atom]         Forces [meV / Bohr]
        Epoch Comm.          Train           Test          Train           Test
--------------------------------------------------------------------------------
RMSE        0     1        38.6582        38.3892       852.3961       841.6591
RMSE        0     2        49.2045        48.4677       843.5274       831.3026
RMSE        0     3        44.9147        43.9323       843.8901       833.0362
RMSE        0     4        45.7725        44.6303       868.3146       859.4484
--------------------------------------------------------------------------------
RMSE        1     1         5.7112         5.1176       272.8571       273.1953
RMSE        1     2         6.3356         8.2201       275.7252       276.8911
RMSE        1     3         5.5069         6.1090       240.5982       239.7464
RMSE        1     4         3.7931         4.0558       230.8009       233.2068
UPDATES     1     1            126                           449
UPDATES     1     2            126                           508
UPDATES     1     3            126                           436
UPDATES     1     4            126                           385
TIMING      1      1.34 min
--------------------------------------------------------------------------------
RMSE        2     1         2.7352         3.1993        93.7605        94.9482
RMSE        2     2         1.8484         1.6521       102.5449       106.3166
RMSE        2     3         2.3983         2.9835        96.2037        95.7769
RMSE        2     4         2.0704         2.0341        95.0648        93.9810
UPDATES     2     1            126                           367
UPDATES     2     2            126                           351
UPDATES     2     3            126                           491
UPDATES     2     4            126                           891
TIMING      2      1.36 min
--------------------------------------------------------------------------------

In each epoch there are now RMSE and UPDATES lines for each committee member, allowing you to monitor the state of the training individually.

All standard output files (e.g., weights_short.XXX.data) are written exactly as in single-network training, but placed in the respective committee subdirectory (1/, 2/, 3/, ...).

Advanced committee options¶

Committee members can be customized individually using the _comm suffix for keywords in input.nn.

Random Seeds per Member¶

You may explicitly define one random seed per committee member:

random_seed 4 8 15 16

The first seed is used to initialize the train / test split.

Architecture Variations¶

The networks architecture can be varied between committee members using

node_short_comm 1 C 10 10
node_short_comm 2 C 15 15

The same can be done for the activations:

activation_nn_comm C 1 t t l
activation_nn_comm C 2 s s l

Fixed parameters¶

Parameters can be fixed and excluded from training. Here as an example for the first committee member.

fix_weights_comm H 1 2 20
fix_biases_comm H 1 2 20

Further documentation can be found in the keyword documentations for fix_weights and fix_biases.

Optimizers¶

Optimizers can be changed between committee members. This is especially useful for fine-tuning of the optimizer settings by training a committee with slightly different optimizer settings.

optimizer 1 1 kalman p_initial=1.0
optimizer 1 2 kalman p_initial=10.0
optimizer 1 3 kalman nue=0.98
optimizer 1 4 kalman nue=0.99

Committee Prediction¶

Committee prediction shares the feature calculation for all committee members. This makes it faster than running all members seperately. For a successful prediction run, all dataset and feature-related files (input.nn, input.data, scaling.data) must be located in your root directory. Committee-specific input files (e.g., the optimized weights.XXX.data files) must be located in their respective 1/, 2/, 3/ folders.

Input¶

As in training, you only need to specify:

num_committee_members 4

If you used additional committee-specific settings during training (like node_comm or activation_nn_comm), reuse those exact keywords in your prediction setup.

Summary

Using committees in RuNNer 2 allows you to:

• Improve robustness of predictions
• Estimate model uncertainty via committee disagreement
• Explore architectural and optimizer diversity
• Perform efficient ensemble prediction with shared feature calculation

Committees are especially recommended for:

• Active learning workflows
• Large-scale screening
• Applications where uncertainty estimation is critical