Facebook recently announced CommAI-env, a platform for training and evaluating an AI system. Inspired by A roadmap towards Machine Intelligence, the system aims for teaching intelligent agents general learning capabilities that would serve as the groundwork for further, more specialized training by human or machine level interaction.
CommAI works by assigning Tasks to the Learner via the Environment and assigning rewards to the Learner for every correct answer. This type of interaction trains the system to identify correct or not answers by the reward. Tasks vary from the simplest silent-not silent variation, to more incrementally complex tasks that require long term memory from the Learner, all applied within the same training session.
This is one but not the only differentiator factor from similar systems like OpenAI Gym or MazeBase. By eliminating the popular division between training and testing sets system aims to generalize Learner’s capabilities on learning new, unknown problems, all the while solving simple, previously seen problems quickly.
The interface for Learner algorithms defines two functions:
next(self, input_bit) serves as the algorithm’s input receiving a bit of input from the environemnt and returning the next bit by the learner.
reward(self, reward) would provide the Learner with the received reward from the environment.
Iteratively working towards rewards will train the algorithm in learning from past data and adapting to new puzzles.
Training platform provides bindings for developers to define learning via Learners in any programming language through zeromq message queues. The project is still in Beta testing phase and provides a Facebook group for discussion and support. The platform is coded in Python and hosted on Github.