The "model" part of a model-based reinforcement learning algorithm.
Pull down the FNC repository. At the command line:
git clone git@codeberg.org:brohrer/cartographer.gitInstall it.
python3 -m pip install -e cartographerThe FNC class Naive Cartographer is intended to be used in tandem with a planner (not included) to build a reinforcement learning agent. At each time step
- The FNC makes a prediction
- The planner uses that prediction to choose an action
- The FNC registers the action selection
The outline of an agent class looks like this:
from cartographer import NaiveCartographer as Model
class Agent
def __init__(self, n_sensors, n_actions):
model = Model(n_sensors=n_sensors, n_actions=n_actions)
planner = ...
def step(self, sensors, reward):
predictions, rewards, uncertainties = model.predict(sensors, reward)
actions = planner.choose_actions(predictions, rewards, uncertainties)
model.update_actions(actions)
return actionsA Fuzzy Naive Cartographer (FNC) builds a Markov Decision Process-like model of its world in the form of a set of sequences. Sequences are of the form of state-action-state as in MDPs. Here they're also referred to as feature-action-outcome sequences, to disambiguate the before- and after-state.
FNC builds a value function too. It creates a set of feature-action pairs and associates a reward with each. This is analogous to the state-action value functions of Q-learning.
A model and a value function allow for prediction and planning. Knowing the current active features and recent actions, both the reward and the resulting features can be anticipated.
Technically FNC is just the "model" part of a model-based RL algorithm. It needs to paired with a planner, some element that will choose an action or goal, to be a complete RL algorithm. There are some rudimentary planners included with this module to get you up and running.
There is more detail in the white paper at https://brandonrohrer.com/cartographer