Julian Togelius

@togelius

AI and games researcher. Associate professor at NYU; Editor-in-Chief of ; director of ; co-founder of .

New York City
Vrijeme pridruživanja: siječanj 2009.
Rođen/a 1979.

Medijski sadržaj

  1. prije 21 sat

    Here's our team, in case you were wondering. We're all standing up straight, or at least perpendicular to some high-dimensional hyperplane that is not depicted because of the inherent limitations of projecting onto a 2D plane.

    Prikaži ovu nit
  2. 4. velj

    Seriously, though. Here's Sutton and Barto (1998), from the introduction to the book Reinforcement Learning. Too many don't understand this.

    Prikaži ovu nit
  3. 1. velj

    It is well known from "non-deep" agent learning research (say, evolutionary robotics work and game AI work from early to mid 2000s) that the sensor representation is extremely important to the agent's ability to learn.

    Prikaži ovu nit
  4. 1. velj

    In our new paper, "Rotation, Translation, and Cropping for Zero-Shot Generalization" by and I show that the agent-centric perspective is better in the sense that the agent learns policies that generalize better.

    Prikaži ovu nit
  5. 1. velj

    To the left, you see a trained agent playing a level of a game. To the right, you see the same playthrough from an agent-centric perspective: cropped, translated, and rotated with the agent in the center. Which perspective is the best input for the agent?

    Prikaži ovu nit
  6. 28. sij

    A really nice piece from about the NYU Game Innovation Lab and the research we do here. Featuring and myself. Thanks for the write-up!

  7. 28. sij

    The difference, instead, is that PCGRL doesn't search for levels. Instead, it searches for level generators. It learns to generate levels. So it moves time consumption from inference time to training time.

    Prikaži ovu nit
  8. 28. sij

    You may wonder how PCGRL relates to search-based PCG (SBPCG). It's not the fact that one approach and the other something. It would be completely possible to do PCGRL with neuroevolution. In this paper we use PPO, but the choice of RL algorithm is not important.

    Prikaži ovu nit
  9. 28. sij

    Finally, the Wide representation differs drastically, in the sense that the network can choose to edit any tile at any timestep. This creates a much larger timestep, but generally much shorter episodes.

    Prikaži ovu nit
  10. 28. sij

    What PCGRL does is that it learns to at any point take the action that maximizes expected future quality. In other words, it learns to improve the level one step at a time. This is particularly interesting for interactive and mixed-initiative approaches.

    Prikaži ovu nit
  11. 28. sij

    The reward function differs between the game. In the maze scenario (previous tweet) it's simply how long the longest path is. In Sokoban, shows here, we want to have as many boxes as possible, as long as the level is solvable. So we include a simple solver in the reward function.

    Prikaži ovu nit
  12. 28. sij

    In the Turtle representation, inspired by Turtle graphics, the neural network can move around freely and edit where it wants. The "agent" has the option to change the tile where it "stands", or move in any cardinal direction. This representation is perhaps the most gamelike one.

    Prikaži ovu nit
  13. 28. sij

    In all cases, we initialize level generation with random levels. This is important to ensure that the network does not simply learn to produce a single level, but instead learns to produce a level generator that can produce an unlimited amount of good new levels.

    Prikaži ovu nit
  14. 28. sij

    We train these networks by presenting a sequence of random positions, but we usually present positions to the network sequentially, like scan lines on an old TV. Interestingly, the network works with almost any position order.

    Prikaži ovu nit
  15. 28. sij

    In the narrow representation, the neural network cannot control where it edits the map; it is presented with a series of positions, and for each of them it needs to decide whether to swap out a particular tile or not.

    Prikaži ovu nit
  16. 28. sij

    What if we see level generation as a game, where each action changes the level in some way? Well, we could use RL to learn to "play" level generation! We introduce Procedural Content Generation via Reinforcement Learning (PCGRL), a new paradigm for PCG.

    Prikaži ovu nit
  17. 16. sij

    我很高兴地向大家宣布:我和 共同撰写的《人工智能与游戏》一书已经被译为中文并出版了!谢谢 的帮助!

    Prikaži ovu nit
  18. 2. sij

    In case I wanted to press any of these buttons (I do not), which would be the more hygienic choice to press? – mjesto: Amsterdam Airport Schiphol (AMS)

  19. 26. pro 2019.

    NB: I'm not writing a novel. I'm barely writing anything at all right now. Here's a road in Skåne in December.

    Prikaži ovu nit
  20. 24. pro 2019.

    Det är alltså jul bland Togeliusarna

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·