Learning

Learning in #bayrob can be facilitated by simply training a JPT model from an existing dataset. The example module provides a convenience configuration for updating/retraining the existing models move, perception, turn and pr2. With minor changes, additional action models can be incorporated into the system. By running python examples/example.py a new action model can be trained, or an existing one can be updated. Each action model has its own folder in examples. The existing models include move, perception, turn and pr2. There are multiple general arguments to control the learning process:

• -v, –verbose {debug,info,warning,error,critical} sets the verbosity level for the current run

• -e, –example <model-name> learns/updates the action model in examples/<model-name>

• -a, –args <arg> <value> passes an action-specific argument to the respective <model-name> module. For accepted arguments, refer to the modules’ documentations

• –recent will address the most recently generated folder (by a previous run). If not given, a new folder with a timestamp of the current run examples/<example-name>/<run-id> is created and the respective module’s generate_data(fp, args) function is called to re-generate the data files, the result of which will be stored in examples/example-name>/<run-id>/data/000-<model-name>.parquet.

• –learn trains JPT models for the <model-name> action. The model object will be stored in examples/example-name>/<run-id>/000-<example-name>.tree

• –modulelearn allows to use a user-specified learning function instead of the default call of learn_jpt in example/example.py

• –crossval triggers the crossvalidation of multiple (predefined) model settings

• –plot will trigger plotting of the action model JPT (twice, with and without variable plots). The result will be stored in examples/example-name>/<run-id>/plots/000-<example-name>.svg and examples/example-name>/<run-id>/plots/000-<example-name>-nodist.svg.

• –min-samples-leaf <n> and –min-impurity-improvement <n> passes the min_samples_leaf or min_impurity_improvement parameter to the learning function

• –obstacles will add obstacle handling in all functions (where necessary)

• –data will trigger generating data/world plots by calling the examples/<model-name> module’s plot_data function

• –prune triggers the do_prune function during the learning of the action model to influence the default behavior

To incorporate a new model into the system, add a new folder <model-name> in examples/ and provide a examples/<model-name>/<model-name>.py module implementing the functions init, generate_data, learn-jpt, plot_data, and tear_down, each accepting two arguments: fp, a string representing the filepath to the current run of the <model-name> and args, an argument object allowing to access the arguments passed to the call.

Once a functioning model has been learnt, it can be integrated into the web application. Place the .tree file of your model along with its .parquet data file in a folder examples/demo/<model-name> which is used by the web application to load the models. The folder structure then looks like the following:

└── examples/
    ├── demo/
    │   ├── alarm
    │   ├── move
    │   ├── ...
    │   └── <model-name>/
    │       ├── 000-<model-name>.tree
    │       └── 000-<model-name>.parquet
    ├── move/
    ├── ...
    ├── <model-name>/
    │   ├── <YYYY-MM-DD_HH:mm>/
    │   │   ├── [crossval/] (optional)
    │   │   ├── data/
    │   │   │   └── 000-<model-name>.parquet
    │   │   ├── plots/
    │   │   │   ├── 000-<model-name>.svg
    │   │   │   ├── 000-<model-name>-nodist.svg
    │   │   │   └── ...
    │   │   └── 000-<model-name>.tree
    │   ├── ...
    │   ├── __init__.py
    │   └── <model-name>.py
    └── ...