Reinforcement Learning in Unity

Reinforcement learning is an artificial intelligence technique that trains agents to perform tasks by rewarding desirable behavior. During reinforcement learning, an agent explores its environment, observes the state of things, and, based on those observations, takes an action. If the action leads to a better state, the agent receives a positive reward. If it leads to a less desirable state, then the agent receives no reward or a negative reward (punishment). As the agent learns during training, it optimizes its decision making so that it receives the maximum reward over time.

The ML-Agents toolkit uses a reinforcement learning technique called Proximal Policy Optimization (PPO). PPO uses a neural network to approximate the ideal function that maps an agent's observations to the best action an agent can take in a given state. The ML-Agents PPO algorithm is implemented in TensorFlow and runs in a separate Python process (communicating with the running Unity application over a socket).

Note: if you aren't studying machine and reinforcement learning as a subject and just want to train agents to accomplish tasks, you can treat PPO training as a black box. There are a few training-related parameters to adjust inside Unity as well as on the Python training side, but you do not need in-depth knowledge of the algorithm itself to successfully create and train agents. Step-by-step procedures for running the training process are provided in the Training section.

The Simulation and Training Process

Training and simulation proceed in steps orchestrated by the ML-Agents Academy class. The Academy works with Agent objects in the scene to step through the simulation. When all Agents in the scene are done, one training episode is finished.

During training, the external Python training process communicates with the Academy to run a series of episodes while it collects data and optimizes its neural network model. When training is completed successfully, you can add the trained model file to your Unity project for later use.

The ML-Agents Academy class orchestrates the agent simulation loop as follows:

Calls your Academy's OnEnvironmentReset delegate.
Calls the AgentReset() function for each Agent in the scene.
Calls the CollectObservations() function for each Agent in the scene.
Uses each Agent's Policy to decide on the Agent's next action.
Calls the AgentAction() function for each Agent in the scene, passing in the action chosen by the Agent's Policy. (This function is not called if the Agent is done.)
Calls the Agent's AgentOnDone() function if the Agent has reached its Max Step count or has otherwise marked itself as done. Optionally, you can set an Agent to restart if it finishes before the end of an episode. In this case, the Academy calls the AgentReset() function.

To create a training environment, extend the Academy and Agent classes to implement the above methods. The Agent.CollectObservations() and Agent.AgentAction() functions are required; the other methods are optional — whether you need to implement them or not depends on your specific scenario.

Note: The API used by the Python training process to communicate with and control the Academy during training can be used for other purposes as well. For example, you could use the API to use Unity as the simulation engine for your own machine learning algorithms. See Python API for more information.

Organizing the Unity Scene

To train and use the ML-Agents toolkit in a Unity scene, the scene must contain a single Academy and as many Agent subclasses as you need. Agent instances should be attached to the GameObject representing that Agent.

Academy

The Academy object orchestrates Agents and their decision making processes. Only place a single Academy object in a scene.

Academy resetting

To alter the environment at the start of each episode, add your method to the Academy's OnEnvironmentReset action.

public class MySceneBehavior : MonoBehaviour
{
    public void Awake()
    {
        var academy = FindObjectOfType<Academy>();
        academy.LazyInitialization();
        academy.OnEnvironmentReset += EnvironmentReset;
    }

    void EnvironmentReset()
    {
        // Reset the scene here
    }
}

For example, you might want to reset an Agent to its starting position or move a goal to a random position. An environment resets when the reset() method is called on the Python UnityEnvironment.

When you reset an environment, consider the factors that should change so that training is generalizable to different conditions. For example, if you were training a maze-solving agent, you would probably want to change the maze itself for each training episode. Otherwise, the agent would probably on learn to solve one, particular maze, not mazes in general.

Agent

The Agent class represents an actor in the scene that collects observations and carries out actions. The Agent class is typically attached to the GameObject in the scene that otherwise represents the actor — for example, to a player object in a football game or a car object in a vehicle simulation. Every Agent must have appropriate Behavior Parameters.

To create an Agent, extend the Agent class and implement the essential CollectObservations() and AgentAction() methods:

CollectObservations() — Collects the Agent's observation of its environment.
AgentAction() — Carries out the action chosen by the Agent's Policy and assigns a reward to the current state.

Your implementations of these functions determine how the Behavior Parameters assigned to this Agent must be set.

You must also determine how an Agent finishes its task or times out. You can manually set an Agent to done in your AgentAction() function when the Agent has finished (or irrevocably failed) its task by calling the Done() function. You can also set the Agent's Max Steps property to a positive value and the Agent will consider itself done after it has taken that many steps. If you set an Agent's ResetOnDone property to true, then the Agent can attempt its task several times in one episode. (Use the Agent.AgentReset() function to prepare the Agent to start again.)

See Agents for detailed information about programming your own Agents.

Environments

An environment in the ML-Agents toolkit can be any scene built in Unity. The Unity scene provides the environment in which agents observe, act, and learn. How you set up the Unity scene to serve as a learning environment really depends on your goal. You may be trying to solve a specific reinforcement learning problem of limited scope, in which case you can use the same scene for both training and for testing trained agents. Or, you may be training agents to operate in a complex game or simulation. In this case, it might be more efficient and practical to create a purpose-built training scene.

Both training and testing (or normal game) scenes must contain an Academy object to control the agent decision making process. When you create a training environment in Unity, you must set up the scene so that it can be controlled by the external training process. Considerations include:

The training scene must start automatically when your Unity application is launched by the training process.
The Academy must reset the scene to a valid starting point for each episode of training.
A training episode must have a definite end — either using Max Steps or by each Agent setting itself to done.

7.6 KiB 原始文件 文件历史