TASSTA Documentation Center TASSTA Documentation Center More products
Hide table of contents Hide details Search My account

Train the location

After you have collected measurements and, optionally, applied the whitelist and blacklist, proceed to the level and location training step. Before you can train a location, you need to have at least one trained level in it. Resolving of mobile user positions in a location starts working only after the location has been trained.

Training for levels and locations is done in T.Rodon.

Training a level

  1. Select the level for which measurements have been collected.
  2. Click the Training button to open the Training menu.
  3. Click Start to initiate the training process. Shortly after that, the progress will be logged in this window. The log entries look like this:
    181/10000 - loss 0,00125 | acc 96,3% | mae 0,02733 | lr 1.0
    Training finishes automatically as soon as sufficiently high estimated accuracy (the acc % value) is reached (for more details, see Configuring Advanced options below).
    Note: Processing increases the load on server resources, so you can process only one level at a time.
  4. When the training finished, the following message is displayed: "Level training is done!".

If you need to interrupt the training process, click Stop.

If the acc % value stays low and stops increasing, try stopping the training and repeating it. If this problem persists, consider redoing your measurements or revising your signal source whitelist.

You can reprocess a level as many times as needed after changes to the measurements, whitelist and blacklist.

To discard the existing training for a level, click the Reset button. Do not confuse this with the Reset that is available during level editing, which deletes collected data.

Training a location

  1. Select a location that has at least one trained level.
  2. Click the Training button to open the Training menu.
  3. Click Start to initiate the training process. Shortly after that, the progress will be logged in this window. The log entries look like this:
    181/10000 - loss 0,00125 | acc 96,3% | mae 0,02733 | lr 1.0
    Training finishes automatically as soon as sufficiently high estimated accuracy (the acc % value) is reached.
    If there is only one trained level in the location, there is usually no log output.
  4. When the training finished, the following message is displayed: "Location training is done!".

If you need to interrupt the training process, click Stop.

If the acc % value stays low and stops increasing, try stopping the training and repeating it. If this problem persists, consider redoing your measurements or revising your signal source whitelist.

You need to reprocess a location whenever you add trained levels to it.

To discard the existing training for a location, click the Reset button. This will not delete the collected measurement data.

Configuring Advanced options

Levels and locations both have Advanced options for training. Change them with caution, because it can greatly increase the load on the server and decrease the resulting accuracy and speed of training instead of increasing them.

The options are as follows:

  • Max Epoch Count
    Number of simulations to do for finding the model with the highest accuracy. The default value is 10000.
    Setting a low value reduces the training time but decreases the resulting accuracy.
    Setting a high value makes the training take longer, but may increase the resulting accuracy if there are large amounts of data.
  • Hidden Layer Count / Hidden Layer Neuron Count
    May require adjustment in some complex situations for higher accuracy. Changing the default value is not recommended.
  • Learning Rate
    Represents how much simulations differ from one another. If set too low, the training process will be too slow. If set too high, simulations may skip the best accuracy.
  • Early Stop Loss
    This value is inversely proportional to accuracy. The lower this value, the better the accuracy. If this is reached before Max Epoch Count, the training process will stop. However, a low Early Stop Loss makes this early stop unlikely.