Arena

Arena is

✅ A building toolkit (code: Arena-BuildingToolkit) for researchers to easily invent and build unexplored single-agent games with tons of research-oriented utilities, it also benefits from a special support to create new multi-agent problems in minutes, see Quick Creation of Social Structures 👉;
✅ A implementation of state-of-the-art deep single/multi-agent reinforcement learning baselines based on RLlib, see code: Arena-Baselines. The supported single-agent baselines are those from RLlib, the supported multi-agent baselines include: independent learners, self-play, population-based training (like in the StarCraft), share weights between arbitrary agents or teams, sharing observations (own observations, team observations, global observations, hidden states), multi-sensors (first/third-person visual, lidar vector), centralized critic and decentralized actor (centralised value baseline) and etc.;
✅ A general evaluation platform for single/multi-agent intelligence, with learning environments of diverse logic and representations, see code: Arena-Benchmark;

♦️ [Research] Research community (especially that of multi-agent) are still at the stage where many problems remain unexplored. Therefore, we provide the Arena-BuildingToolkit for researchers to easily invent and build novel multi-agent problems, as well as easily test the new problem with the existing Arena-Baselines. For this part, built upon ML-Agent, we 1) make things more compact and easy, and 2) have a special support for multi-agent.
♦️ [Application] Considering the pipeline of applying a acting agent (say, an RL agent) to a specific real-world task, you would need to: 1) build simulation environment, 2) code an agent interface to interact with it, 3) train the agent and check if the agent is making sensible actions, and 4) deploy the agent. Our vision is to make step 1-3 easily accessible. Thus, one of our future plans would be making the simulation as realistic as possible and having metric to evaluate how realistic the simulation is. For this part, it is overlapped with the engine behind this project ML-Agent.

Arena-AAAI20-Spotlight-Presentation

Jan. 2020: Arena-Baselines has been re-created based on RLlib with more baselines and better utilities. The old one is depreciated and left in Arena-Baselines-Depreciated.
Oct. 2019: Paper accepted in AAAI-2020, see a lot of you in New York!
Jul. 2019: Open Slack workspace for discussion, joint us on Slack.
Jun. 2019: Open Trello board for Arena-Baselines and Trello board for Arena-BuildingToolkit, contributions are welcomed.
May. 2019: Code release on Github, Arena-Baselines and Arena-BuildingToolkit.