Before I start, I just want to mention how DIFFICULT it was to set up the ML-agents python package. I had to install different an older version of python, use a specific version of different things etc. I don’t think I have ever had such a headache just to install a python package, jesus!
Anyways, When I don’t run the CMD as admin it does a PermissionError: [WinError 5] Access is denied, so I run it as a admin. But when i finally got to training my agents, i got the error: ValueError: semaphore or lock released too many times.
I can’t fix this. I’m using python v 3.7.2, MLAgents package is 0.29.0, mlagentenvs 0.29.0.
Not sure why it should make a huge difference but here is my c# agent code (runs fine in heuristic):
using UnityEngine;
using Unity.MLAgents;
using Unity.MLAgents.Sensors;
using Unity.MLAgents.Actuators;
using System;
public enum Team
{
Red,
Blue
}
public class Bean : Agent
{
[SerializeField] private float moveSpeed = 5f;
private CharacterController characterController;
private Vector3 moveDirection;
[SerializeField] private Team team;
[SerializeField] private Transform TeamMate;
[SerializeField] private Transform OpposingBean1;
[SerializeField] private Transform OpposingBean2;
[SerializeField] private Transform ball;
[SerializeField] private Transform goal1;
[SerializeField] private Transform goal2;
[SerializeField] private Vector3 AgentStartPos;
private void Start()
{
characterController = GetComponent<CharacterController>();
Debug.Log("Start position: " + ball.transform.position);
}
public override void OnEpisodeBegin()
{
Debug.Log("Episode started.");
transform.position = AgentStartPos;
ball.position = new Vector3(0, 1, 0);
}
// Collect observations from the environment
public override void CollectObservations(VectorSensor sensor)
{
Debug.Log("Collecting observations.");
sensor.AddObservation(transform.position);
sensor.AddObservation(TeamMate.transform.position);
sensor.AddObservation(OpposingBean1.transform.position);
sensor.AddObservation(OpposingBean2.transform.position);
sensor.AddObservation(ball.transform.position);
sensor.AddObservation(goal1.transform.position);
sensor.AddObservation(goal2.transform.position);
}
// Process actions received from the agent
public override void OnActionReceived(ActionBuffers actions)
{
float horizontalMovement = actions.ContinuousActions[0];
float verticalMovement = actions.ContinuousActions[1];
moveDirection = transform.TransformDirection(new Vector3(horizontalMovement, 0f, verticalMovement).normalized);
// Move the agent using the CharacterController (ensure characterController is not null)
if (characterController != null)
{
characterController.SimpleMove(moveDirection * moveSpeed);
Debug.Log("Moving: " + moveDirection);
}
else
{
Debug.LogWarning("CharacterController is null. Make sure it's assigned in the Inspector.");
}
}
public void AddGoalReward(Team rewardTeam)
{
if (rewardTeam == Team.Blue)
{
if (team == Team.Blue)
{
AddReward(2f);
EndEpisode();
Debug.Log("Blue team scored. Reward: +2");
}
else
{
AddReward(-2f);
EndEpisode();
Debug.Log("Red team scored. Reward: -2");
}
}
else if (rewardTeam == Team.Red)
{
if (team == Team.Red)
{
AddReward(2f);
EndEpisode();
Debug.Log("Red team scored. Reward: +2");
}
else
{
AddReward(-2f);
EndEpisode();
Debug.Log("Blue team scored. Reward: -2");
}
}
}
// Heuristic method for manual control (using arrow keys)
public override void Heuristic(in ActionBuffers actionsOut)
{
ActionSegment<float> continuousActions = actionsOut.ContinuousActions;
float horizontalInput = Input.GetAxis("Horizontal");
float verticalInput = Input.GetAxis("Vertical");
continuousActions[0] = horizontalInput;
continuousActions[1] = verticalInput;
Debug.Log("Heuristic - Horizontal: " + horizontalInput + ", Vertical: " + verticalInput);
}
}
Thanks in advance.