This repository is a private approach to training a DQN agent to solve nonogram puzzles.
(see https://en.wikipedia.org/wiki/Nonogram)
While the environment itself was created, the DQN and agent rely heavily on this repository:
https://github.com/sourcecode369/deep-reinforcement-learning/tree/master/DQN/LunarLander-v0
Thanks a lot!
State type: Discrete(64)
Number of actions: 50
More information will follow soon.
With the given code, 3 training runs with 500 epochs each were performed.
The results are quite promising, but not yet final.
For evaluation, the trained agents were each given a random puzzle.
They have 100 steps to solve this puzzle.
In the current setup, a state is reached after 10-15 steps in which the agent is stuck / only performs one action in a continuous loop. (@FutureWork)
As can be seen in the following three images, mistakes are also still made that may prevent the goal.
During the training the following values are achieved:
Episode 100 Average Score: -8958.15
Episode 200 Average Score: -2564.90
Episode 300 Average Score: -2467.80
Episode 400 Average Score: -2642.08
Episode 500 Average Score: -2180.86
Episode 100 Average Score: -8642.55
Episode 200 Average Score: -2401.74
Episode 300 Average Score: -2151.25
Episode 400 Average Score: -1892.52
Episode 500 Average Score: -1214.48
Episode 100 Average Score: -8527.56
Episode 200 Average Score: -3132.81
Episode 300 Average Score: -2688.90
Episode 400 Average Score: -2788.20
Episode 500 Average Score: -2608.07
