Our new paper, by @Amidos2006 @FilipoGiovanni @Smearle_RH and myself, lays out the conceptual framework for learning level generators with reinforcement learning and provides an initial Deep RL implementation.
https://arxiv.org/abs/2001.09212
-
-
Prikaži ovu nit
-
We use three different game scenarios, Sokoban, the GVGAI version of Zelda, and a simple maze framework. The reward function and available tiles differ for each. We also introduce three different representations: narrow, turtle, and wide.
Prikaži ovu nit -
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.pic.twitter.com/oBuS796rq9
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
Interesting! Just skimming the paper: is the criteria for a good level, whether or not it is solvable by an agent? I would guess we would want levels that were "goldilocks" (not too hard not too easy) in difficulty to provide a good curriculum .
-
Not sure whether this has been the standard in the literature, but POET had such a not too hard and not too easy scheme.
- Još 5 drugih odgovora
Novi razgovor -
-
-
Awesome work. Looking forward to reading in detail. I had a paper from AIIDE-2016 where I used vanilla MCTS-based planning to generate Sokoban levels from simulated game-play. http://motion.cs.umn.edu/pub/SokobanMCTS/DataDrivenSokobanMCTS.pdf …
-
Oh nice! Thanks! This actually ties even better in to another paper which we are working on...
@Amidos2006@Bumblebor - Još 2 druga odgovora
Novi razgovor -
-
-
Could this be combined within an adversarial/cooperative setup where one agent creates levels and another plays them and they both get better at it? Probably rather difficult to stabilize because it's not transitive but could be neat

-
Yes. The basic idea of having a learning agent in the loop for game/level generation was proposed some time ago: https://ieeexplore.ieee.org/abstract/document/5035629/ … We're actually currently working on a related approach, where we train a network to both generate levels and play games.
Kraj razgovora
Novi razgovor -
-
-
Evaluating levels remains the hard part and using 'solvers' is the logical solution here, but I genuinely would like to see human trials to cross-reference the generated levels with player feedback. Wonder if there's a simple game with lots of UGC that would fit (~Geometry dash)pic.twitter.com/pYvRP1sSJ0
-
That could be a great future work :) as we tried to focus in this paper on introducing the problem itself and ideas of how to transform level generation to markov decision process :)
- Još 7 drugih odgovora
Novi razgovor -
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.