I am requesting research, articles, abstracts or interesting opinions that will help me create a complex causal neural network. There are many detailed resources on causal discovery, image recognition, spatial invariance. However, none of them take into account an animated changing game state, and all are very specific and hard to use for complex problems. Here is the core of the problem:
I am making a 2D strategy simulation/game and would like to create a neural network that can learn and deduce all the game's rules. The game, similar to Conway's Game of Life, takes the form of a 2-dimensional image, with red green and blue values for each pixel/cell. The game goes in turns, where every turn, the pixels of the image are evaluated, and changed based on the rules dictated by their RGB values. The network will train on a large data set of animated images, representing the changes to the game's state, with each frame being a turn.
RULES PER PIXEL
- if red is above 50%, all adjacent pixels have their blue values halved.
- if blue is below 75%, between 2-8 random pixels within a 10 tile radius will increase their green values by 25%
- if green value is divisible by 5, all adjacent pixels will turn white (100%red, 100% blue, 100% green)
- if pixel is black, 15% of the map's pixels will be turned white at random locations.
- if pixel is white, a random pixel on the image will have their RGB values randomized.
- if green value is above 80%, the first 3 pixels to the immediate north will all have their values switched with each other.
The rules are very arbitrary and feature a mix of conditional checks as well as random changes to the map. This is intended to challenge the network's spatial and causal recognition. There is also the intention of adding or removing rules in the future, with the network being flexible enough to learn the new rules or changes to old rules, if found.
The main constraint I placed is that there will be no feature engineering, and the game's rules will always be hidden from the network. The challenge is for this theoretical network to analyze short and long term changes to the game state, discover and deduce as much of the game's hidden rules as possible, in the time it's given to analyze the data. It will eventually need to design its own generated image, designed in order to maximize the lifetime of an arbitrary choice of color. For example, it may be tasked to keep as much of the image red as possible, or possibly keep the green values low. The idea is that giving it new design constraints should be intuitive and easy, without having to re-code or re-train the entire network every time. I would also like to avoid using pooling, sampling, or any other tricks for image-based neural networks to detect patterns. This is because some of the game's rules are arbitrarily random, and may not have direct spatial correlation, while other rules do.