A reinforcement learning algorithm for the 2048 game using decision trees and a genetic algorithm.
The AI doesn't know the consequences of it's actions, it only see the game state and it's score. No heuristic is used : the goal is to see if it is possible for an AI to learn how to solve the game by itself.
The DNA contains k decision trees.
- A mutation consists in replace one of the trees by an other random one.
- A crossover consists in picking k random trees from each parent to build a new DNA
Each decision tree has a fixed height. Each node compares a random feature with a random value choosed in the value-range of the selected feature. Leafs contains either a decision either an "unknown decision".
The decision taken by an individual is choosed by a majority vote of each decision tree. Of course, unknown decisions are not considered in the vote.
My vision of this design is that each tree "can be seen as a detector" of some specific configurations of the board that should lead to a specific action.
It could be a good idea to process a meta-optimisation of these parameters :
- Population size
- Decision trees height
- Mutation rate
- Ratio of decision/unknown decision in the leafs of decision trees
At the moment, the only extracted features is an log histogram of the board. It could be a good idea to count the mergeable tiles for each direction, variance, or other features.
Finally, may be try to design a "real" genetic programming algorithm :D
These results has been measured at the git rev 214d1a68b9b3053a3dc44ca446cf80974d071032
At the moment, it is able to learn but it has not solved the game yet ! Best score : 10620 (after 10min training, did a 1024 tile)
Other interesting trainig sessions :
