ml-agents/docs/ML-Agents-Overview.md

ML-Agents Toolkit Overview

The Unity Machine Learning Agents Toolkit (ML-Agents Toolkit) is an open-source project that enables games and simulations to serve as environments for training intelligent agents. Agents can be trained using reinforcement learning, imitation learning, neuroevolution, or other machine learning methods through a simple-to-use Python API. We also provide implementations (based on TensorFlow) of state-of-the-art algorithms to enable game developers and hobbyists to easily train intelligent agents for 2D, 3D and VR/AR games. These trained agents can be used for multiple purposes, including controlling NPC behavior (in a variety of settings such as multi-agent and adversarial), automated testing of game builds and evaluating different game design decisions pre-release. The ML-Agents toolkit is mutually beneficial for both game developers and AI researchers as it provides a central platform where advances in AI can be evaluated on Unity’s rich environments and then made accessible to the wider research and game developer communities.

Depending on your background (i.e. researcher, game developer, hobbyist), you may have very different questions on your mind at the moment. To make your transition to the ML-Agents toolkit easier, we provide several background pages that include overviews and helpful resources on the Unity Engine, machine learning and TensorFlow. We strongly recommend browsing the relevant background pages if you're not familiar with a Unity scene, basic machine learning concepts or have not previously heard of TensorFlow.

The remainder of this page contains a deep dive into ML-Agents, its key components, different training modes and scenarios. By the end of it, you should have a good sense of what the ML-Agents toolkit allows you to do. The subsequent documentation pages provide examples of how to use ML-Agents. To get started, watch this demo video of ML-Agents in action.

Running Example: Training NPC Behaviors

To help explain the material and terminology in this page, we'll use a hypothetical, running example throughout. We will explore the problem of training the behavior of a non-playable character (NPC) in a game. (An NPC is a game character that is never controlled by a human player and its behavior is pre-defined by the game developer.) More specifically, let's assume we're building a multi-player, war-themed game in which players control the soldiers. In this game, we have a single NPC who serves as a medic, finding and reviving wounded players. Lastly, let us assume that there are two teams, each with five players and one NPC medic.

The behavior of a medic is quite complex. It first needs to avoid getting injured, which requires detecting when it is in danger and moving to a safe location. Second, it needs to be aware of which of its team members are injured and require assistance. In the case of multiple injuries, it needs to assess the degree of injury and decide who to help first. Lastly, a good medic will always place itself in a position where it can quickly help its team members. Factoring in all of these traits means that at every instance, the medic needs to measure several attributes of the environment (e.g. position of team members, position of enemies, which of its team members are injured and to what degree) and then decide on an action (e.g. hide from enemy fire, move to help one of its members). Given the large number of settings of the environment and the large number of actions that the medic can take, defining and implementing such complex behaviors by hand is challenging and prone to errors.

With ML-Agents, it is possible to train the behaviors of such NPCs (called agents) using a variety of methods. The basic idea is quite simple. We need to define three entities at every moment of the game (called environment):

• Observations - what the medic perceives about the environment. Observations can be numeric and/or visual. Numeric observations measure attributes of the environment from the point of view of the agent. For our medic this would be attributes of the battlefield that are visible to it. For most interesting environments, an agent will require several continuous numeric observations. Visual observations, on the other hand, are images generated from the cameras attached to the agent and represent what the agent is seeing at that point in time. It is common to confuse an agent's observation with the environment (or game) state. The environment state represents information about the entire scene containing all the game characters. The agents observation, however, only contains information that the agent is aware of and is typically a subset of the environment state. For example, the medic observation cannot include information about an enemy in hiding that the medic is unaware of.
• Actions - what actions the medic can take. Similar to observations, actions can either be continuous or discrete depending on the complexity of the environment and agent. In the case of the medic, if the environment is a simple grid world where only their location matters, then a discrete action taking on one of four values (north, south, east, west) suffices. However, if the environment is more complex and the medic can move freely then using two continuous actions (one for direction and another for speed) is more appropriate.
• Reward signals - a scalar value indicating how well the medic is doing. Note that the reward signal need not be provided at every moment, but only when the medic performs an action that is good or bad. For example, it can receive a large negative reward if it dies, a modest positive reward whenever it revives a wounded team member, and a modest negative reward when a wounded team member dies due to lack of assistance. Note that the reward signal is how the objectives of the task are communicated to the agent, so they need to be set up in a manner where maximizing reward generates the desired optimal behavior.

After defining these three entities (the building blocks of a reinforcement learning task), we can now train the medic's behavior. This is achieved by simulating the environment for many trials where the medic, over time, learns what is the optimal action to take for every observation it measures by maximizing its future reward. The key is that by learning the actions that maximize its reward, the medic is learning the behaviors that make it a good medic (i.e. one who saves the most number of lives). In reinforcement learning terminology, the behavior that is learned is called a policy, which is essentially a (optimal) mapping from observations to actions. Note that the process of learning a policy through running simulations is called the training phase, while playing the game with an NPC that is using its learned policy is called the inference phase.

The ML-Agents toolkit provides all the necessary tools for using Unity as the simulation engine for learning the policies of different objects in a Unity environment. In the next few sections, we discuss how the ML-Agents toolkit achieves this and what features it provides.

Key Components

The ML-Agents Toolkit contains four high-level components:

• Learning Environment - which contains the Unity scene and all the game characters. 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.
• Python API - which contains a low-level Python interface for interacting and manipulating a learning environment. Note that, unlike the Learning Environment, the Python API is not part of Unity, but lives outside and communicates with Unity through the Communicator. This API is contained in a dedicated mlagents_envs Python package and is used by the Python training process to communicate with and control the Academy during training. However, it 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.
• Communicator - which connects the Learning Environment with the Python API. It lives within the Learning Environment.
• Python Trainers which contains all the machine learning algorithms that enable training agents. The algorithms are implemented in Python and are part of their own mlagents Python package. The package exposes a single command-line utility mlagents-learn that supports all the training methods and options outlined in this document. The Python Trainers interface solely with the Python API.

Example block diagram of ML-Agents toolkit for our sample game.

The Learning Environment contains an additional component that help organize the Unity scene:

• Agents - which is attached to a Unity GameObject (any character within a scene) and handles generating its observations, performing the actions it receives and assigning a reward (positive / negative) when appropriate. Each Agent is linked to a Policy.

Every Learning Environment will always have one Agent for every character in the scene. While each Agent must be linked to a Policy, it is possible for Agents that have similar observations and actions to have the same Policy type. In our sample game, we have two teams each with their own medic. Thus we will have two Agents in our Learning Environment, one for each medic, but both of these medics can have the same Policy. Note that these two medics have the same Policy because their space of observations and actions are similar. This does not mean that at each instance they will have identical observation and action values. In other words, the Policy defines the space of all possible observations and actions, while the Agents connected to it (in this case the medics) can each have their own, unique observation and action values. If we expanded our game to include tank driver NPCs, then the Agent attached to those characters cannot share a Policy with the Agent linked to the medics (medics and drivers have different actions).

We have yet to discuss how the ML-Agents toolkit trains behaviors, and what role the Python API and External Communicator play. Before we dive into those details, let's summarize the earlier components. Each character is attached to an Agent, and each Agent has a Policy. The Policy receives observations and rewards from the Agent and returns actions. The Academy ensures that all the Agents are in sync in addition to controlling environment-wide settings.

Training Modes

Given the flexibility of ML-Agents, there are a few ways in which training and inference can proceed.

Built-in Training and Inference

As mentioned previously, the ML-Agents toolkit ships with several implementations of state-of-the-art algorithms for training intelligent agents. More specifically, during training, all the medics in the scene send their observations to the Python API through the External Communicator. The Python API processes these observations and sends back actions for each medic to take. During training these actions are mostly exploratory to help the Python API learn the best policy for each medic. Once training concludes, the learned policy for each medic can be exported. Given that all our implementations are based on TensorFlow, the learned policy is just a TensorFlow model file. Then during the inference phase, we use the TensorFlow model generated from the training phase. Now during the inference phase, the medics still continue to generate their observations, but instead of being sent to the Python API, they will be fed into their (internal, embedded) model to generate the optimal action for each medic to take at every point in time.

To summarize: our built-in implementations are based on TensorFlow, thus, during training the Python API uses the observations it receives to learn a TensorFlow model. This model is then embedded within the Agent during inference.

The Getting Started Guide tutorial covers this training mode with the 3D Balance Ball sample environment.

Custom Training and Inference

In the previous mode, the Agents were used for training to generate a TensorFlow model that the Agents can later use. However, any user of the ML-Agents toolkit can leverage their own algorithms for training. In this case, the behaviors of all the Agents in the scene will be controlled within Python. You can even turn your environment into a gym.

We do not currently have a tutorial highlighting this mode, but you can learn more about the Python API here.

Flexible Training Scenarios

While the discussion so-far has mostly focused on training a single agent, with ML-Agents, several training scenarios are possible. We are excited to see what kinds of novel and fun environments the community creates. For those new to training intelligent agents, below are a few examples that can serve as inspiration:

• Single-Agent. A single agent, with its own reward signal. The traditional way of training an agent. An example is any single-player game, such as Chicken.
• Simultaneous Single-Agent. Multiple independent agents with independent reward signals with same Behavior Parameters. A parallelized version of the traditional training scenario, which can speed-up and stabilize the training process. Helpful when you have multiple versions of the same character in an environment who should learn similar behaviors. An example might be training a dozen robot-arms to each open a door simultaneously.
• Adversarial Self-Play. Two interacting agents with inverse reward signals. In two-player games, adversarial self-play can allow an agent to become increasingly more skilled, while always having the perfectly matched opponent: itself. This was the strategy employed when training AlphaGo, and more recently used by OpenAI to train a human-beating 1-vs-1 Dota 2 agent.
• Cooperative Multi-Agent. Multiple interacting agents with a shared reward signal with same or different Behavior Parameters. In this scenario, all agents must work together to accomplish a task that cannot be done alone. Examples include environments where each agent only has access to partial information, which needs to be shared in order to accomplish the task or collaboratively solve a puzzle.
• Competitive Multi-Agent. Multiple interacting agents with inverse reward signals with same or different Behavior Parameters. In this scenario, agents must compete with one another to either win a competition, or obtain some limited set of resources. All team sports fall into this scenario.
• Ecosystem. Multiple interacting agents with independent reward signals with same or different Behavior Parameters. This scenario can be thought of as creating a small world in which animals with different goals all interact, such as a savanna in which there might be zebras, elephants and giraffes, or an autonomous driving simulation within an urban environment.

Training Methods

This section overviews the various state-of-the-art machine learning algorithms that are part of the ML-Agents Toolkit. If you aren't studying machine and reinforcement learning as a subject and just want to train agents to accomplish tasks, you can treat these algorithms as black boxes. 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 algorithms themselves to successfully create and train agents. Step-by-step procedures for running the training process are provided in the Training ML-Agents page.

A Quick Note on Reward Signals

In this section we introduce the concepts of intrinsic and extrinsic rewards, which helps explain some of the training methods.

To train an agent, you will need to provide the agent one or more reward signals which the agent should attempt to maximize. In reinforcement learning, the end goal for the Agent is to discover a behavior (a Policy) that maximizes a reward. Typically, a reward is defined by your environment, and corresponds to reaching some goal. These are what we refer to as "extrinsic" rewards, as they are defined external of the learning algorithm.

Rewards, however, can be defined outside of the environment as well, to encourage the agent to behave in certain ways, or to aid the learning of the true extrinsic reward. We refer to these rewards as "intrinsic" reward signals. The total reward that the agent will learn to maximize can be a mix of extrinsic and intrinsic reward signals.

The ML-Agents Toolkit allows reward signals to be defined in a modular way, and we provide three reward signals that can the mixed and matched to help shape your agent's behavior:

• extrinsic: represents the rewards defined in your environment, and is enabled by default
• gail: represents an intrinsic reward signal that is defined by GAIL (see below)
• curiosity: represents an intrinsic reward signal that encourages exploration in sparse-reward environments that is defined by the Curiosity module (see below)

Deep Reinforcement Learning

ML-Agents provide an implementation of two reinforcement learning algorithms:

• Proximal Policy Optimization (PPO)
• Soft Actor-Critic (SAC)

The default algorithm is PPO. This is a method that has been shown to be more general purpose and stable than many other RL algorithms.

In contrast with PPO, SAC is off-policy, which means it can learn from experiences collected at any time during the past. As experiences are collected, they are placed in an experience replay buffer and randomly drawn during training. This makes SAC significantly more sample-efficient, often requiring 5-10 times less samples to learn the same task as PPO. However, SAC tends to require more model updates. SAC is a good choice for heavier or slower environments (about 0.1 seconds per step or more). SAC is also a "maximum entropy" algorithm, and enables exploration in an intrinsic way. Read more about maximum entropy RL here.

Curiosity for Sparse-reward Environments

In environments where the agent receives rare or infrequent rewards (i.e. sparse-reward), an agent may never receive a reward signal on which to bootstrap its training process. This is a scenario where the use of an intrinsic reward signals can be valuable. Curiosity is one such signal which can help the agent explore when extrinsic rewards are sparse.

The curiosity Reward Signal enables the Intrinsic Curiosity Module. This is an implementation of the approach described in [Curiosity-driven Exploration by Self-supervised Prediction(https://arxiv.org/abs/1705.05363). It trains two networks:

• an inverse model, which takes the current and next observation of the agent, encodes them, and uses the encoding to predict the action that was taken between the observations
• a forward model, which takes the encoded current observation and action, and predicts the next encoded observation.

The loss of the forward model (the difference between the predicted and actual encoded observations) is used as the intrinsic reward, so the more surprised the model is, the larger the reward will be.

For more information, see our dedicated blog post on the Curiosity module.

Imitation Learning

It is often more intuitive to simply demonstrate the behavior we want an agent to perform, rather than attempting to have it learn via trial-and-error methods. For example, instead of indirectly training a medic with the help of a reward function, we can give the medic real world examples of observations from the game and actions from a game controller to guide the medic's behavior. Imitation Learning uses pairs of observations and actions from a demonstration to learn a policy. See this video demo of imitation learning .

Imitation learning can either be used alone or in conjunction with reinforcement learning. If used alone it can provide a mechanism for learning a specific type of behavior (i.e. a specific style of solving the task). If used in conjunction with reinforcement learning it can dramatically reduce the time the agent takes to solve the environment. This can be especially pronounced in sparse-reward environments. For instance, on the Pyramids environment, using 6 episodes of demonstrations can reduce training steps by more than 4 times. See Behavioral Cloning + GAIL + Curiosity + RL below.

The ML-Agents Toolkit provides a way to learn directly from demonstrations, as well as use them to help speed up reward-based training (RL). We include two algorithms called Behavioral Cloning (BC) and Generative Adversarial Imitation Learning (GAIL). In most scenarios, you can combine these two features.

GAIL (Generative Adversarial Imitation Learning)

GAIL, or Generative Adversarial Imitation Learning, is an imitation learning algorithm that uses an adversarial approach, in a similar vein to GANs (Generative Adversarial Networks). In this framework, a second neural network, the discriminator, is taught to distinguish whether an observation/action is from a demonstration or produced by the agent. This discriminator can the examine a new observation/action and provide it a reward based on how close it believes this new observation/action is to the provided demonstrations.

At each training step, the agent tries to learn how to maximize this reward. Then, the discriminator is trained to better distinguish between demonstrations and agent state/actions. In this way, while the agent gets better and better at mimicking the demonstrations, the discriminator keeps getting stricter and stricter and the agent must try harder to "fool" it.

This approach learns a policy that produces states and actions similar to the demonstrations, requiring fewer demonstrations than direct cloning of the actions. In addition to learning purely from demonstrations, the GAIL reward signal can be mixed with an extrinsic reward signal to guide the learning process.

GAIL uses an adversarial approach to reward your Agent for behaving similar to a set of demonstrations. GAIL can be used with or without environment rewards, and works well when there are a limited number of demonstrations.

Behavioral Cloning (BC)

BC trains the Agent's policy to exactly mimic the actions shown in a set of demonstrations. The BC feature can be enabled on the PPO or SAC trainers. As BC cannot generalize past the examples shown in the demonstrations, BC tends to work best when there exists demonstrations for nearly all of the states that the agent can experience, or in conjunction with GAIL and/or an extrinsic reward.

Recording Demonstrations

Demonstrations of agent behavior can be recorded from the Unity Editor or build, and saved as assets. These demonstrations contain information on the observations, actions, and rewards for a given agent during the recording session. They can be managed in the Editor, as well as used for training with BC and GAIL.

Summary

To summarize, we provide 3 training methods: BC, GAIL and RL (PPO or SAC) that can be used independently or together:

• BC can be used on its own or as a pre-training step before GAIL and/or RL
• GAIL can be used with or without extrinsic rewards
• RL can be used on its own (either PPO or SAC) or in conjunction with BC and/or GAIL.

Leveraging either BC or GAIL requires recording demonstrations to be provided as input to the training algorithms.

Training Options

In addition to the three training methods introduced in the previous section, the ML-Agents Toolkit provides additional options that can aid in training behaviors in complex environments. Each of the options below can be utilized with any of the training methods above.

Training in Multi-Agent Environments with Self-Play

ML-Agents provides the functionality to train both symmetric and asymmetric adversarial games with Self-Play. A symmetric game is one in which opposing agents are equal in form, function and objective. Examples of symmetric games are our Tennis and Soccer example environments. In reinforcement learning, this means both agents have the same observation and action spaces and learn from the same reward function and so they can share the same policy. In asymmetric games, this is not the case. An example of an asymmetric games are Hide and Seek. Agents in these types of games do not always have the same observation or action spaces and so sharing policy networks is not necessarily ideal.

With self-play, an agent learns in adversarial games by competing against fixed, past versions of its opponent (which could be itself as in symmetric games) to provide a more stable, stationary learning environment. This is compared to competing against the current, best opponent in every episode, which is constantly changing (because it's learning).

Self-play can be used with our implementations of both Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). However, from the perspective of an individual agent, these scenarios appear to have non-stationary dynamics because the opponent is often changing. This can cause significant issues in the experience replay mechanism used by SAC. Thus, we recommend that users use PPO. For further reading on this issue in particular, see the paper Stabilising Experience Replay for Deep Multi-Agent Reinforcement Learning.

Memory-enhanced Agents using Recurrent Neural Networks

Have you ever entered a room to get something and immediately forgot what you were looking for? Don't let that happen to your agents.

In some scenarios, agents must learn to remember the past in order to take the best decision. When an agent only has partial observability of the environment, keeping track of past observations can help the agent learn. Deciding what the agents should remember in order to solve a task is not easy to do by hand, but our training algorithms can learn to keep track of what is important to remember with LSTM.

Solving Complex Tasks using Curriculum Learning

Curriculum learning is a way of training a machine learning model where more difficult aspects of a problem are gradually introduced in such a way that the model is always optimally challenged. This idea has been around for a long time, and it is how we humans typically learn. If you imagine any childhood primary school education, there is an ordering of classes and topics. Arithmetic is taught before algebra, for example. Likewise, algebra is taught before calculus. The skills and knowledge learned in the earlier subjects provide a scaffolding for later lessons. The same principle can be applied to machine learning, where training on easier tasks can provide a scaffolding for harder tasks in the future.

Imagine training the medic to to scale a wall to arrive at a wounded team member. The starting point when training a medic to accomplish this task will be a random policy. That starting policy will have the medic running in circles, and will likely never, or very rarely scale the wall properly to revive their team member (and achieve the reward). If we start with a simpler task, such as moving toward an unobstructed team member, then the medic can easily learn to accomplish the task. From there, we can slowly add to the difficulty of the task by increasing the size of the wall until the medic can complete the initially near-impossible task of scaling the wall. We have included an environment to demonstrate this with ML-Agents, called Wall Jump.

Demonstration of a curriculum training scenario in which a progressively taller wall obstructs the path to the goal.

To see curriculum learning in action, observe the two learning curves below. Each displays the reward over time for an agent trained using PPO with the same set of training hyperparameters. The difference is that one agent was trained using the full-height wall version of the task, and the other agent was trained using the curriculum version of the task. As you can see, without using curriculum learning the agent has a lot of difficulty. We think that by using well-crafted curricula, agents trained using reinforcement learning will be able to accomplish tasks otherwise much more difficult.

The ML-Agents toolkit supports setting custom environment parameters within the Academy. This allows elements of the environment related to difficulty or complexity to be dynamically adjusted based on training progress.

Training Robust Agents using Environment Parameter Randomization

An agent trained on a specific environment, may be unable to generalize to any tweaks or variations in the environment (in machine learning this is referred to as overfitting). This becomes problematic in cases where environments are instantiated with varying objects or properties. One mechanism to alleviate this and train more robust agents that can generalize to unseen variations of the environment is to expose them to these variations during training. Similar to Curriculum Learning, where environments become more difficult as the agent learns, the ML-Agents Toolkit provides a way to randomly sample parameters of the environment during training. We refer to this approach as Environment Parameter Randomization. For those familiar with Reinforcement Learning research, this approach is based on the concept of Domain Randomization (you can read more about it here). By using parameter randomization during training, the agent can be better suited to adapt (with higher performance) to future unseen variations of the environment.

Example of variations of the 3D Ball environment.

Ball scale of 0.5	Ball scale of 4

Additional Features

Beyond the flexible training scenarios available, the ML-Agents Toolkit includes additional features which improve the flexibility and interpretability of the training process.

• Monitoring Agent’s Decision Making - Since communication in ML-Agents is a two-way street, we provide an Agent Monitor class in Unity which can display aspects of the trained Agent, such as the Agents perception on how well it is doing (called value estimates) within the Unity environment itself. By leveraging Unity as a visualization tool and providing these outputs in real-time, researchers and developers can more easily debug an Agent’s behavior. You can learn more about using the Monitor class here.

• Complex Visual Observations - Unlike other platforms, where the agent’s observation might be limited to a single vector or image, the ML-Agents toolkit allows multiple cameras to be used for observations per agent. This enables agents to learn to integrate information from multiple visual streams. This can be helpful in several scenarios such as training a self-driving car which requires multiple cameras with different viewpoints, or a navigational agent which might need to integrate aerial and first-person visuals. You can learn more about adding visual observations to an agent here.

• Concurrent Unity Instances - We enable developers to run concurrent, parallel instances of the Unity executable during training. For certain scenarios, this should speed up training.

Summary and Next Steps

To briefly summarize: The ML-Agents toolkit enables games and simulations built in Unity to serve as the platform for training intelligent agents. It is designed to enable a large variety of training modes and scenarios and comes packed with several features to enable researchers and developers to leverage (and enhance) machine learning within Unity.

To help you use ML-Agents, we've created several in-depth tutorials for installing ML-Agents, getting started with the 3D Balance Ball environment (one of our many sample environments) and making your own environment.