Trying to solve the simple chain environment using LSPI about mushroom-rl HOT 1 CLOSED

This is not a bug, the behavior is expected. LSPI automatically computes the action features from the state features. You just need to set use input_size=1, as the state vector has only one component.

Remember also that for the simple chain there is no visualization, so you have to turn off the rendering.

However, you are not supposed to use the polynomial features for finite states, as they are specifically designed for vectors of continuous state-actions.
In our framework the index of the discrete state is not meant to have a meaning, it is just a label for the state. Of course, in a chain, this label is meaningful as it represents the order in the chain.
In general, if you really want to create features for finite-state MDPs, you can just create your own class, implementing the FeaturesImplementation (in features._implementations) interface. This will let you choose a reasonable set of features to the MDP, if you have previous information about the state distribution.

I hope this solves your issue.

from mushroom-rl.