Skip to content

Analysis Tools

Jump to bottom
Marco Marini edited this page May 2, 2026 · 8 revisions

No longer valid since release 0.25.0

Octave tools

The analysis tools are located in octave folder. To collect the kpis the -k <kpi_folder> command option is required during agent training.

For information about performance analysis please refer to Analysis of Performance

Agent performance

The octave program td_report generates the agent performance report. The kpi folder and the report folder should be selected.

Delta weights

To collect the statistics of the weight variations it is necessary to run the training session by generating all the kpis with the command options -k <kpi_folder> -l all.

The weight changes are ploted with td_delta_weights.m octave program.

Motor theta corrections

The speed of the motors depends on the voltage supplied. The dependence is non-linear mainly due to static friction. For this reason, the supplied voltage is a non-linear function of the power signal which corrects for these non-linearities. The parameters $(\theta)$ of this function can be deduced by measuring the speed of the motors with respect to the supplied voltage (see MotorMeasuring wheellino project).

The Octave motorMeasureReport.m application allows you to calculate the theta parameters in speed measurements and than can be transpose in the robot configuration yaml file.

Weights statistics

To dump neural network weight statistics, after installed the application run the command:

>java -classpath lib/wheelly-0.6.0.jar org.mmarini.wheelly.apps.WeightsStats

the options are

>java -classpath lib/wheelly-0.6.0.jar org.mmarini.wheelly.apps.WeightsStats -h

usage: org.mmarini.wheelly.apps.WeightsStats
       [-h] [-v] modelPath

Dump the weights statistic

positional arguments:
  modelPath              specify the model path

named arguments:
  -h, --help             show this help message and exit
  -v, --version          show current version

The number of weights, the minimum and maximum weights, the number of biases, the minimum and maximum values ​​of bias for the critic and actor network will be displayed.

>java -classpath lib/wheelly-0.6.0.jar org.mmarini.wheelly.apps.WeightsStats

Model: models/polarRadarAnalogAgent

Policy
  41800 weights
  Min value: -0,0357621
  Max value: 0,0292231
  442 bias
  Min value: -0,174762
  Max value: 0,353298

Critic
  37700 weights
  Min value: -0,0215887
  Max value: 0,0245415
  401 bias
  Min value: -0,00191194
  Max value: 7,14404e-05

Release 0.5.2

Clone this wiki locally