Work
- Implemented abstract class for experiments; easy to setup an experiment now
- Implemented some re-usable print outs experiments
- Implemented method to output an array to a csv-file for easy evaluation of results
- Implemented basic structure regarding MCTS, Regression trees and TLS
- Implemented a simple UCT agent discretizing the action space
- Parameter: discretization steps (number of actions per action dimension)
- Parameter: Exploration constant (the "C" in the UCT formula)
- Did more reading and some initial coding regarding VFDT / SDR / Hoeffding / FIMT
Problems Encountered
- Due to observations, nodes in the tree can represent only estimates of the states, i.e. Cart Pole has different starting states meaning the root of the MCTS tree does not always represent the same starting state.
- I do not yet fully understand VFDT / FIMT
- How are split points defined in case of a continuous range? Papers only state: "for each possible split point"?
- What is an example in this context?
- Continue on incremental regression tree induction
No comments:
Post a Comment