Unity 机器学习代理工具包 (ML-Agents) 是一个开源项目，它使游戏和模拟能够作为训练智能代理的环境。

unity3d unity unity-tech reinforcement-le deep-learning deep-reinforcement-learning neural-networks

21 KiB

原始文件 文件历史

Making a New Learning Environment

This tutorial walks through the process of creating a Unity Environment. A Unity Environment is an application built using the Unity Engine which can be used to train Reinforcement Learning Agents.

In this example, we will train a ball to roll to a randomly placed cube. The ball also learns to avoid falling off the platform.

Overview

Using the ML-Agents toolkit in a Unity project involves the following basic steps:

Create an environment for your agents to live in. An environment can range from a simple physical simulation containing a few objects to an entire game or ecosystem.
Add an Academy MonoBehaviour to a GameObject in the Unity scene containing the environment.
Implement your Agent subclasses. An Agent subclass defines the code an Agent uses to observe its environment, to carry out assigned actions, and to calculate the rewards used for reinforcement training. You can also implement optional methods to reset the Agent when it has finished or failed its task.
Add your Agent subclasses to appropriate GameObjects, typically, the object in the scene that represents the Agent in the simulation.

Note: If you are unfamiliar with Unity, refer to Learning the interface in the Unity Manual if an Editor task isn't explained sufficiently in this tutorial.

If you haven't already, follow the installation instructions.

Set Up the Unity Project

The first task to accomplish is simply creating a new Unity project and importing the ML-Agents assets into it:

Launch the Unity Editor and create a new project named "RollerBall".
Make sure that the Scripting Runtime Version for the project is set to use .NET 4.x Equivalent (This is an experimental option in Unity 2017, but is the default as of 2018.3.)
In a file system window, navigate to the folder containing your cloned ML-Agents repository.
Drag the ML-Agents folder from UnitySDK/Assets to the Unity Editor Project window. If you see console errors about Barracuda, make sure you've installed Barracuda from the Unity Package Manager. More information can be found in the installation instructions under Package Installation.

Your Unity Project window should contain the following assets:

Create the Environment

Next, we will create a very simple scene to act as our ML-Agents environment. The "physical" components of the environment include a Plane to act as the floor for the Agent to move around on, a Cube to act as the goal or target for the agent to seek, and a Sphere to represent the Agent itself.

Create the Floor Plane

Right click in Hierarchy window, select 3D Object > Plane.
Name the GameObject "Floor."
Select the Floor Plane to view its properties in the Inspector window.
Set Transform to Position = (0, 0, 0), Rotation = (0, 0, 0), Scale = (1, 1, 1).
On the Plane's Mesh Renderer, expand the Materials property and change the default-material to LightGridFloorSquare (or any suitable material of your choice).

(To set a new material, click the small circle icon next to the current material name. This opens the Object Picker dialog so that you can choose a different material from the list of all materials currently in the project.)

Add the Target Cube

Right click in Hierarchy window, select 3D Object > Cube.
Name the GameObject "Target"
Select the Target Cube to view its properties in the Inspector window.
Set Transform to Position = (3, 0.5, 3), Rotation = (0, 0, 0), Scale = (1, 1, 1).
On the Cube's Mesh Renderer, expand the Materials property and change the default-material to Block.

Add the Agent Sphere

Right click in Hierarchy window, select 3D Object > Sphere.
Name the GameObject "RollerAgent"
Select the RollerAgent Sphere to view its properties in the Inspector window.
Set Transform to Position = (0, 0.5, 0), Rotation = (0, 0, 0), Scale = (1, 1, 1).
On the Sphere's Mesh Renderer, expand the Materials property and change the default-material to CheckerSquare.
Click Add Component.
Add the Physics/Rigidbody component to the Sphere.

Note that we will create an Agent subclass to add to this GameObject as a component later in the tutorial.

Add an Empty GameObject to Hold the Academy

Right click in Hierarchy window, select Create Empty.
Name the GameObject "Academy"

You can adjust the camera angles to give a better view of the scene at runtime. The next steps will be to create and add the ML-Agent components.

Add an Academy

The Academy object coordinates the ML-Agents in the scene and drives the decision-making portion of the simulation loop. Every ML-Agent scene needs one (and only one) Academy instance.

First, add an Academy component to the Academy GameObject created earlier:

Select the Academy GameObject to view it in the Inspector window.
Click Add Component.
Select Academy in the list of components.

Implement an Agent

To create the Agent:

Select the RollerAgent GameObject to view it in the Inspector window.
Click Add Component.
Click New Script in the list of components (at the bottom).
Name the script "RollerAgent".
Click Create and Add.

Then, edit the new RollerAgent script:

In the Unity Project window, double-click the RollerAgent script to open it in your code editor.
In the editor, add the using MLAgents; statement and then change the base class from MonoBehaviour to Agent.
Delete the Update() method, but we will use the Start() function, so leave it alone for now.

So far, these are the basic steps that you would use to add ML-Agents to any Unity project. Next, we will add the logic that will let our Agent learn to roll to the cube using reinforcement learning.

Initialization and Resetting the Agent

When the Agent reaches its target, it marks itself done and its Agent reset function moves the target to a random location. In addition, if the Agent rolls off the platform, the reset function puts it back onto the floor.

To move the target GameObject, we need a reference to its Transform (which stores a GameObject's position, orientation and scale in the 3D world). To get this reference, add a public field of type Transform to the RollerAgent class. Public fields of a component in Unity get displayed in the Inspector window, allowing you to choose which GameObject to use as the target in the Unity Editor.

To reset the Agent's velocity (and later to apply force to move the agent) we need a reference to the Rigidbody component. A Rigidbody is Unity's primary element for physics simulation. (See Physics for full documentation of Unity physics.) Since the Rigidbody component is on the same GameObject as our Agent script, the best way to get this reference is using GameObject.GetComponent<T>(), which we can call in our script's Start() method.

So far, our RollerAgent script looks like:

using System.Collections.Generic;
using UnityEngine;
using MLAgents;

public class RollerAgent : Agent
{
    Rigidbody rBody;
    void Start () {
        rBody = GetComponent<Rigidbody>();
    }

    public Transform Target;
    public override void AgentReset()
    {
        if (this.transform.position.y < 0)
        {
            // If the Agent fell, zero its momentum
            this.rBody.angularVelocity = Vector3.zero;
            this.rBody.velocity = Vector3.zero;
            this.transform.position = new Vector3( 0, 0.5f, 0);
        }

        // Move the target to a new spot
        Target.position = new Vector3(Random.value * 8 - 4,
                                      0.5f,
                                      Random.value * 8 - 4);
    }
}

Next, let's implement the Agent.CollectObservations() method.

Observing the Environment

The Agent sends the information we collect to the Brain, which uses it to make a decision. When you train the Agent (or use a trained model), the data is fed into a neural network as a feature vector. For an Agent to successfully learn a task, we need to provide the correct information. A good rule of thumb for deciding what information to collect is to consider what you would need to calculate an analytical solution to the problem.

In our case, the information our Agent collects includes:

Position of the target.

AddVectorObs(Target.position);

Position of the Agent itself.

AddVectorObs(this.transform.position);

The velocity of the Agent. This helps the Agent learn to control its speed so it doesn't overshoot the target and roll off the platform.

// Agent velocity
AddVectorObs(rBody.velocity.x);
AddVectorObs(rBody.velocity.z);

In total, the state observation contains 8 values and we need to use the continuous state space when we get around to setting the Brain properties:

public override void CollectObservations()
{
    // Target and Agent positions
    AddVectorObs(Target.position);
    AddVectorObs(this.transform.position);

    // Agent velocity
    AddVectorObs(rBody.velocity.x);
    AddVectorObs(rBody.velocity.z);
}

The final part of the Agent code is the Agent.AgentAction() method, which receives the decision from the Brain and assigns the reward.

Actions

The decision of the Brain comes in the form of an action array passed to the AgentAction() function. The number of elements in this array is determined by the Vector Action Space Type and Space Size settings of the agent's Brain. The RollerAgent uses the continuous vector action space and needs two continuous control signals from the Brain. Thus, we will set the Brain Space Size to 2. The first element,action[0] determines the force applied along the x axis; action[1] determines the force applied along the z axis. (If we allowed the Agent to move in three dimensions, then we would need to set Vector Action Size to 3.) Note that the Brain really has no idea what the values in the action array mean. The training process just adjusts the action values in response to the observation input and then sees what kind of rewards it gets as a result.

The RollerAgent applies the values from the action[] array to its Rigidbody component, rBody, using the Rigidbody.AddForce function:

Vector3 controlSignal = Vector3.zero;
controlSignal.x = action[0];
controlSignal.z = action[1];
rBody.AddForce(controlSignal * speed);

Rewards

Reinforcement learning requires rewards. Assign rewards in the AgentAction() function. The learning algorithm uses the rewards assigned to the Agent during the simulation and learning process to determine whether it is giving the Agent the optimal actions. You want to reward an Agent for completing the assigned task. In this case, the Agent is given a reward of 1.0 for reaching the Target cube.

The RollerAgent calculates the distance to detect when it reaches the target. When it does, the code calls the Agent.SetReward() method to assign a reward of 1.0 and marks the agent as finished by calling the Done() method on the Agent.

float distanceToTarget = Vector3.Distance(this.transform.position,
                                          Target.position);
// Reached target
if (distanceToTarget < 1.42f)
{
    SetReward(1.0f);
    Done();
}

Note: When you mark an Agent as done, it stops its activity until it is reset. You can have the Agent reset immediately, by setting the Agent.ResetOnDone property to true in the inspector or you can wait for the Academy to reset the environment. This RollerBall environment relies on the ResetOnDone mechanism and doesn't set a Max Steps limit for the Academy (so it never resets the environment).

Finally, if the Agent falls off the platform, set the Agent to done so that it can reset itself:

// Fell off platform
if (this.transform.position.y < 0)
{
    Done();
}

AgentAction()

With the action and reward logic outlined above, the final version of the AgentAction() function looks like:

public float speed = 10;
public override void AgentAction(float[] vectorAction)
{
    // Actions, size = 2
    Vector3 controlSignal = Vector3.zero;
    controlSignal.x = vectorAction[0];
    controlSignal.z = vectorAction[1];
    rBody.AddForce(controlSignal * speed);

    // Rewards
    float distanceToTarget = Vector3.Distance(this.transform.position,
                                              Target.position);

    // Reached target
    if (distanceToTarget < 1.42f)
    {
        SetReward(1.0f);
        Done();
    }

    // Fell off platform
    if (this.transform.position.y < 0)
    {
        Done();
    }

}

Note the speed class variable defined before the function. Since speed is public, you can set the value from the Inspector window.

Final Editor Setup

Now, that all the GameObjects and ML-Agent components are in place, it is time to connect everything together in the Unity Editor. This involves changing some of the Agent Component's properties so that they are compatible with our Agent code.

Select the RollerAgent GameObject to show its properties in the Inspector window.
Change Decision Interval from 1 to 10.
Drag the Target GameObject from the Hierarchy window to the RollerAgent Target field.
Add the Behavior Parameters script with the Add Component button from the RollerAgent Inspector.
Modify the Behavior Parameters of the Agent :

Behavior Name to RollerBallBrain
Vector Observation Space Size = 8
Vector Action Space Type = Continuous
Vector Action Space Size = 2

Now you are ready to test the environment before training.

Testing the Environment

It is always a good idea to test your environment manually before embarking on an extended training run. To do so, you will need to implement the Heuristic() method on the RollerAgent class. This will allow you control the Agent using direct keyboard control.

The Heuristic() method will look like this :

    public override float[] Heuristic()
    {
        var action = new float[2];
        action[0] = Input.GetAxis("Horizontal");
        action[1] = Input.GetAxis("Vertical");
        return action;
    }

What this code means is that the heuristic will generate an action corresponding to the values of the "Horizontal" and "Vertical" input axis (which correspond to the keyboard arrow keys).

In order for the Agent to use the Heuristic, You will need to check the Use Heuristic checkbox in the Behavior Parameters of the RollerAgent.

Press Play to run the scene and use the arrows keys to move the Agent around the platform. Make sure that there are no errors displayed in the Unity editor Console window and that the Agent resets when it reaches its target or falls from the platform. Note that for more involved debugging, the ML-Agents SDK includes a convenient Monitor class that you can use to easily display Agent status information in the Game window.

One additional test you can perform is to first ensure that your environment and the Python API work as expected using the notebooks/getting-started.ipynb Jupyter notebook. Within the notebook, be sure to set env_name to the name of the environment file you specify when building this environment.

Training the Environment

The process is the same as described in Training ML-Agents. Note that the models will be created in the original ml-agents project folder, ml-agents/models.

The hyperparameters for training are specified in the configuration file that you pass to the mlagents-learn program. Using the default settings specified in the original ml-agents/config/trainer_config.yaml file, the RollerAgent takes about 300,000 steps to train. However, you can change the following hyperparameters to speed up training considerably (to under 20,000 steps):

batch_size: 10
buffer_size: 100

Since this example creates a very simple training environment with only a few inputs and outputs, using small batch and buffer sizes speeds up the training considerably. However, if you add more complexity to the environment or change the reward or observation functions, you might also find that training performs better with different hyperparameter values.

Note: In addition to setting these hyperparameter values, the Agent DecisionFrequency parameter has a large effect on training time and success. A larger value reduces the number of decisions the training algorithm has to consider and, in this simple environment, speeds up training.

To train in the editor, run the following Python command from a Terminal or Console window before pressing play:

mlagents-learn config/config.yaml --run-id=RollerBall-1 --train

(where config.yaml is a copy of trainer_config.yaml that you have edited to change the batch_size and buffer_size hyperparameters for your trainer.)

Note: If you get a command not found error when running this command, make sure that you have followed the Install Python and mlagents Package section of the ML-Agents Installation instructions.

To monitor the statistics of Agent performance during training, use TensorBoard.

In particular, the cumulative_reward and value_estimate statistics show how well the Agent is achieving the task. In this example, the maximum reward an Agent can earn is 1.0, so these statistics approach that value when the Agent has successfully solved the problem.

Note: If you use TensorBoard, always increment or change the run-id you pass to the mlagents-learn command for each training run. If you use the same id value, the statistics for multiple runs are combined and become difficult to interpret.

Optional: Multiple Training Areas within the Same Scene

In many of the example environments, many copies of the training area are instantiated in the scene. This generally speeds up training, allowing the environment to gather many experiences in parallel. This can be achieved simply by instantiating many Agents which share the Behavior Parameters. Use the following steps to parallelize your RollerBall environment.

Instantiating Multiple Training Areas

Right-click on your Project Hierarchy and create a new empty GameObject. Name it TrainingArea.
Reset the TrainingArea’s Transform so that it is at (0,0,0) with Rotation (0,0,0) and Scale (1,1,1).
Drag the Floor, Target, and RollerAgent GameObjects in the Hierarchy into the TrainingArea GameObject.
Drag the TrainingArea GameObject, along with its attached GameObjects, into your Assets browser, turning it into a prefab.
You can now instantiate copies of the TrainingArea prefab. Drag them into your scene, positioning them so that they do not overlap.

Editing the Scripts

You will notice that in the previous section, we wrote our scripts assuming that our TrainingArea was at (0,0,0), performing checks such as this.transform.position.y < 0 to determine whether our agent has fallen off the platform. We will need to change this if we are to use multiple TrainingAreas throughout the scene.

A quick way to adapt our current code is to use localPosition rather than position, so that our position reference is in reference to the prefab TrainingArea's location, and not global coordinates.

Replace all references of this.transform.position in RollerAgent.cs with this.transform.localPosition.
Replace all references of Target.position in RollerAgent.cs with Target.localPosition.

This is only one way to achieve this objective. Refer to the example environments for other ways we can achieve relative positioning.

Review: Scene Layout

This section briefly reviews how to organize your scene when using Agents in your Unity environment.

There are two kinds of game objects you need to include in your scene in order to use Unity ML-Agents: an Academy and one or more Agents.

Keep in mind:

There can only be one Academy game object in a scene.
If you are using multiple training areas, make sure all the Agents have the same Behavior Name and Behavior Parameters

21 KiB 原始文件 文件历史