2048 resources
| url | commennt |
|---|---|
| https://arxiv.org/abs/1604.05085 | Best known AI |
| http://www.cs.put.poznan.pl/wjaskowski/pub/papers/Szubert2014_2048.pdf | the first base improved by above |
| https://github.com/kennedymj97/2048-AI | rust engine |
| http://www.mit.edu/~amarj/files/2048.pdf | a paper from "MIT" that failed to get very far with basic RL approaches to 2048. Perhaps it's written by undergrad students. Not sure But their experience mirrors mine with simple approaches to 2048 |
| https://github.com/navjindervirdee/2048-deep-reinforcement-learning | Played 2048 can achieve 10% of the time which is on par with playing each direction 10 times and taking the best |
| https://github.com/voice32/2048_RL | Article This guy listed everything he's tried; many of the things he's tried is similar to my work |
| https://www.linkedin.com/pulse/part-1-applied-reinforcement-learning-2048-tianyi-pan/ | Interesting score function which is 10 biggest cell + sum of all cells) to give a big boost to cells |
| https://cs.uwaterloo.ca/~mli/zalevine-dqn-2048.pdf | Got deep learning to play better than the author ; but still not super-human level play |
| https://tjwei.github.io/2048-NN/ | Achieves 2048 at 94% probability with a 9 layer fully connect network. Seems like essentially a policy graident approach; seems fake, as the 9 layer neural network runs pretty fast on the internet. So not sure. |
| [Reinforcement Learning For Constraint Satisfaction Game Agents (15-Puzzle, Minesweeper, 2048, and Sudoku)(https://arxiv.org/abs/2102.06019) | 40% to 2048 |
| https://github.com/gorgitko/MI-MVI_2016 | another failed attempt. just used dense network; not even CNN |
| Awesome stackoverflow posts | One post by Xiao is regarded as the best open source implementation |
Is it possible to build an agent that can play to super human levels WITHOUT search?