google-research / falken

Falken provides developers with a service that allows them to train AI that can play their games
Apache License 2.0
253 stars 34 forks source link
cpp games imitation-learning ml unity3d

Falken

Main branch Latest release

NOTE: This is not an officially supported Google product.

Falken provides developers with a service that allows them to train AI that can play their games. Unlike traditional RL frameworks that learn through rewards or batches of offline training, Falken is based on training AI via realtime, human interactions.

Because human time is precious, the system is designed to seamlessly transition between humans playing the game (training) and the AI playing the game (inference). This allows us to support a variety of training styles, from providing a large batch of demonstrations up front to providing “just in time” demonstrations when the AI encounters a state it does not know how to handle.

Behind the scenes, Falken automatically handles logging of experience, separates out human demonstrations, trains new models, serves the best models for on-device inference, and maintains a version history.

Components

This project consists of:

Getting started

The fastest way to ensure that everything is working properly is to launch a local service then build and run Hello Falken environment.

In C++:

In Unity:

Concepts

Stand up and Initialize Falken

The first step in using Falken is to launch the service and connect to the service from the client SDK (e.g C++ or Unity) using your JSON configuration. All subsequent API calls from the client will automatically use this information.

Define Actions

Actions allow you to define how Falken’s AI interacts with your game and how it learns from humans who play it. In Falken, Actions can be either continuous or discrete.

Define Observations

Observations allow you to define how Falken’s AI perceives your game both while it is playing and while it is “watching” a human play. Falken’s Observations are based on the Entity/Attribute model, where an Entity is simply a container for attributes, each of which have a corresponding name, type, and value.

Defining the best Observations for your game can be tricky and can impact Falken's performance, so we recommend starting with the bare minimum a human would require to play a game and slowly refining the set of Observations to improve efficacy. In practice we've found that simple observations can yield much better performance.

Create a Brain

In Falken, a Brain represents the idea of the thing that observes, learns, and acts. It’s also helpful to think about each Brain as being designed to accomplish a specific kind of task, like completing a lap in a racing game or defeating enemies while moving through a level in an FPS.

When creating a Brain, you provide a specific set of Actions and a specific set of Observations that define how the brain perceives the world and what actions it can take.

Start a Session. Start an Episode.

In Falken, a Session represents a period of time during which Falken was learning how to play your game, playing your game, or both. Sessions are composed of one or more Episodes, which represent a sequence of interactions with a clear beginning, middle, and end (like starting a race, driving around the track, and crossing the finish line).

Falken supports a number of different Session types, including:

All Sessions also require you to define a maximum number of Steps per Episode. You’ll want to set this number to be a comfortable amount more than one a human would require. For example, if you call Episode::Step() every frame, your game runs at 30fps, and a human would take 300 seconds to complete an episode, then you should set a max steps to about 9000 or 12000 to ensure the AI has enough time to complete the task. If the number of steps is too high then Falken will take longer to detect when an agent is stuck when using an evaluation session.

Step the Episode

Every time you want Falken to learn from a human or interact with your game, you need to need to call Step(). Step() always operates on a set of Observations from your game. If a human is playing, then you also provide the Actions that the human performed, and Falken will learn from them. If you want Falken to play your game, don’t pass in any Actions to Step() but instead use the Actions returned by Step() and apply them to your game as if a human had performed them.

For most games, you’ll call Falken’s Step() every time you update your game simulation. However, some games might want to call Step() less often, like when the player takes a turn or performs an important action.

End the Episode. Reset the Game.

Episodes in Falken end when the player succeeds, fails, or the maximum number of steps has been exceeded. You provide this context when you end the Episode.

Most Sessions involve many Episodes (either for training Falken or testing your game), so when an Episode completes, you’ll need to reset your game state before you bring the next Episode. Falken leaves the definition of this reset up to you and does not require that every Episode in a Session start with identical state. In general, you want each episode to start with similar but not identical state (similar to what many games do for human players) as this helps Falken most effectively learn from and test your game.

Ending one Session. Starting a new Session.

Once you are done with a Session, you simply Stop it. If an Episode is still active when a Session is stopped, the Episode will be Aborted.

When you Stop a learning-type session (InteractiveTraining or Evaluation), the Brain makes a record of what it has learned and "gets smarter". Additional Sessions created with this brain will start from this more intelligent state and use that as the basis for additional learning or playing.