using System.Collections.Generic;
using Unity.MLAgents.Actuators;
using Unity.MLAgents.Analytics;
using Unity.MLAgents.Sensors;
namespace Unity.MLAgents.Policies
{
///
/// The Remote Policy only works when training.
/// When training your Agents, the RemotePolicy will be controlled by Python.
///
internal class RemotePolicy : IPolicy
{
int m_AgentId;
string m_FullyQualifiedBehaviorName;
ActionSpec m_ActionSpec;
ActionBuffers m_LastActionBuffer;
private bool m_AnalyticsSent = false;
internal ICommunicator m_Communicator;
///
/// List of actuators, only used for analytics
///
private IList m_Actuators;
///
public RemotePolicy(
ActionSpec actionSpec,
IList actuators,
string fullyQualifiedBehaviorName)
{
m_FullyQualifiedBehaviorName = fullyQualifiedBehaviorName;
m_Communicator = Academy.Instance.Communicator;
m_Communicator?.SubscribeBrain(m_FullyQualifiedBehaviorName, actionSpec);
m_ActionSpec = actionSpec;
m_Actuators = actuators;
}
///
public void RequestDecision(AgentInfo info, List sensors)
{
if (!m_AnalyticsSent)
{
m_AnalyticsSent = true;
TrainingAnalytics.RemotePolicyInitialized(
m_FullyQualifiedBehaviorName,
sensors,
m_ActionSpec,
m_Actuators
);
}
m_AgentId = info.episodeId;
m_Communicator?.PutObservations(m_FullyQualifiedBehaviorName, info, sensors);
}
///
public ref readonly ActionBuffers DecideAction()
{
m_Communicator?.DecideBatch();
var actions = m_Communicator?.GetActions(m_FullyQualifiedBehaviorName, m_AgentId);
m_LastActionBuffer = actions == null ? ActionBuffers.Empty : (ActionBuffers)actions;
return ref m_LastActionBuffer;
}
public void Dispose()
{
}
}
}