Class ObservableAttribute

Specify that a field or property should be used to generate observations for an Agent. For each field or property that uses ObservableAttribute, a corresponding ISensor will be created during Agent initialization, and this sensor will read the values during training and inference.

Inheritance

Object

ObservableAttribute

Namespace: Unity.MLAgents.Sensors.Reflection

Syntax

public class ObservableAttribute : Attribute

Remarks

ObservableAttribute is intended to make initial setup of an Agent easier. Because it uses reflection to read the values of fields and properties at runtime, this may be much slower than reading the values directly. If the performance of ObservableAttribute is an issue, you can get the same functionality by overriding CollectObservations(VectorSensor) or creating a custom ISensor implementation to read the values without reflection.

Note that you do not need to adjust the VectorObservationSize in BrainParameters when adding ObservableAttribute to fields or properties.

Examples

This sample class will produce two observations, one for the m_Health field, and one for the HealthPercent property.

using Unity.MLAgents;
using Unity.MLAgents.Sensors.Reflection;

public class MyAgent : Agent
{
    [Observable]
    int m_Health;

    [Observable]
    float HealthPercent
    {
        get => return 100.0f * m_Health / float(m_MaxHealth);
    }
}

Constructors

ObservableAttribute(String, Int32)

ObservableAttribute constructor.

Declaration

public ObservableAttribute(string name = null, int numStackedObservations = 1)

Parameters

Type	Name	Description
String	name	Optional override for the sensor name. Note that all sensors for an Agent must have a unique name.
Int32	numStackedObservations	Number of frames to concatenate observations from.