Consider the feature-based reinforcement learner at: http://artint.info/demos/rl/sGameFA.html. In this question you are going to play with the set of features. You can change the set of features in SGameFeatureSet.java; to remove a feature just comment it out (and it will have value 0).
Give a minimal subset of the existing features for which the agent can eventually learn a policy with a positive average reward. Explain what each of the features does.
What happens when redundant features are added?
What happens if some features are multiplied by a constant?
What happens if some features are negative?
Does it perform better if the same or product of existing features are added?
Give two (new or existing) features that when added to your minimimal set of features gives the most improvement together. Bonus marks will be given to the student(s) who find features with the best performance.