Feedback 15-03-12

General Structures for Nodes and Trees

• I've implemented some general structures for  nodes and trees
• These structures are based on the shared interfaces
  

Observation Tree

• I've implemented (and tested) an incremental regression tree learner
• It's based on the FIMT (see references from previous post) and TG [1,2] algorithms
• Input examples are instances of type Observation from RL-Glue
• Handles only one-dimensional observations at the moment
• Uses the F-Test for finding splits
• Critical values for the F-distribution from 1 to 1000 are stored in an array
• There are three parameters to influence the splitting:
1. The minimum number of examples which need to be seen before the algorithm tests for possible splits
2. The maximum number of split points that can be stored in a node
3. The minimum number of examples which need to be seen for each "side" of a split point before the algorithm considers this split point
• For split points the first n examples are taken (where n is equal to the 2nd parameter described above)

Planning

• Implement simulator for the Tiger Problem
• Combine regression tree with MCTS

[1] Driessens, K., Ramon, J., & Blockeel, H. (2001). Speeding up relational reinforcement learning through the use of an incremental first order decision tree learner. In L. De Raedt & P. Flach (Eds.), Machine Learning ECML 2001 (Vol. 2167, pp. 97-108). Springer.
[2]  Cristina Hohan Yu Chang. Relational Reinforcement Learning with the Algorithm TG. Politecnico di Milano. 2010.