Why I Read This
Arthur Juliani, the first author of this paper, announced the Obstacle Tower Challenge in his Twitter 4 days ago (on January 29, 2019). I enjoy participating in reinforcement learning competitions, since they help me learn difficult ideas while having fun. (In fact, I learned about DQNs through the OpenAI Retro Contest!)
I read Unity’s short post about the challenge, and found out that they had already released the environment on GitHub. I played the game myself, and also took a brief look at the environment. I was excited, since it was much more challenging than existing discrete-action-space environments. The competition begins on February 11th, but I decided to read the paper in advance to gain a better understanding of the environment.
A good benchmark is essential for rapid progress in field. ImageNet, a large image database, had a significant impact on the progress of computer vision research since its presentation in 2009. Similarly the Arcade Learning Environment (ALE) and OpenAI Gym enabled rapid research in reinforcement learning. ALE offers Atari 2600 games as reinforcement learning environments, and OpenAI Gym offers a unified interface for agent-environment interaction.
AI benchmarks lose a lot of its purpose once the AI’s performance exceeds that of humans. If the AI has a subhuman performance, humans can analyze the mistakes to understand how the AI can be improved. However, if the AI has a superhuman performance, it becomes much more challenging to understand how to improve the performance further.
New reinforcement learning agents have shown that they have superhuman performance in most games in ALE. R2D2 by DeepMind have superhuman performance in 52 of 57 Atari games with the same architecture. The five remaining games are Montezuma’s Revenge, Pitfall, Private Eye, Skiing, and Solaris. Go-Explore by Uber achieved superhuman performance on Montezuma’s Revenge and Pitfall. Thus, the authors believe that a more challenging environment is necessary to foster progress in reinforcement learning.
Obstacle Tower conists of up to 100 floors (0-99), with the player starting on floor zero. Each floor is a single finite episode, and it may contain puzzles to solve, enemies to defeat, obstacles to evade, or keys to unlock doors. As the agent progresses to higher floors, the floor layout becomes more complex. The episode terminates when the player encounters a hazard (pit, enemy), is out of time, or finishes the game by reaching the top floor. There are spase and dense reward functions: sparse reward function only give +1 reward on clearing the floor, while dense reward function also gives +0.1 reward on opening doors and clearing puzzles.
The authors compare the Obstacle Tower environment with previous reinforcement learning environments such as ALE, VizDoom, Gym Retro, CoinRun, and Pommerman. The authors differentiate Obstacle Tower from these environments by listing these features:
Physics-driven interactions The player’s interaction with objects are controlled by a 3D physics engine (Unity).
High visual fidelity The game is much more photo-realistic than retro games like ALE, VizDoom, or Gym Retro.
Procedural generation of nontrivial levels and puzzles Unlike most environments with fixed map, Obstacle Tower procedurally generates levels and puzzles, testing the generalization capabiilities of the agent.
Procedurally varied graphics Obstacle Tower also tests the agent’s visual generalization capabilities by varying the texture, lighting, and geometry of the objects.
The authors also list what challenges the agent provides:
Generalization In this stochastic environment, both the floor plan and the room layout. Thus, simple imitation learning methods or methods that take advantage of the environment’s determinism will perform poorly. Instead, the agent must be able to generalize well.
Vision The main source of observation is a $168 \times 168$ rendered RGB image with different textures, lighting, and complex shapes. The authors believe that agents must have a “greater representational capacity” for this environment.
Control The Obstacle Tower includes obstacles, enemies, and moving platforms, which require fine-tuned control.
Planning Floor plans get complex as floors get higher, so the agent must plan the path appropriately.
Popular baseline algorithms - Rainbow and Proximal Policy Optimization (PPO) - perform poorly on a subhuman level. The authors end by suggesting a few subfields of reinforcement learning algorithms that could prove to be most likely improve the performance of the agent:
Hierarchical Control There is a natural hierarchy - floor and room - in this environment.
Intrinsic Motivation Even with the dense reward function, the environment has sparse rewards relative to most existing reinforcement learning environments. In fact, the dense reward function is very similar to that of Montezuma’s Revenge. Thus, augmenting the reward function could lead to faster learning.
Meta-Learning The agent must learn to quickly adapt to different variations of environments.
World-Model Learning Since the environment is based on 3D physics, learning a relatively high-level movement dynamics could be easier for the agent rather than learning a complex control sequence.
- The Obstacle Tower environment can perhaps be better summarized as a 3D stochastic version of Montezuma’s Revenge with an easy version of Sokoban.
- The environment sounds perhaps too difficult: it seems like it requires an agent with good exploration and planning, paired with a good convolutional neural network (CNN).
If you want to learn more about the Arcade Learning Environment (ALE), the predecessor of Obstacle Tower environment, check these links.
- Arcade Learning Environment (GitHub Repo)
- The Arcade Learning Environment: An Evaluation Platform for General Agents (arXiv Paper)
- Revisiting the Arcade Learning Environment: Evaluation Protocols and Open Problems for General Agents (arXiv Paper)
If you want to learn more about the environment, check these links.