The ModelBasedTabularOffPE class extends TabularOffPE using a model-based approach. It relies on empirical estimates of initial distributions, transition dynamics, and rewards to perform off-policy evaluation in tabular settings, by estimating

• d0 the initial state-action distribution $d_0^\pi$,
• dD the state-action dataset distribution $d^D$,
• P the transition matrix $P^\pi$, and
• r the expected rewards $r$.

def __init__(self, dataset, n_obs, n_act, path=None, verbosity=0, auxiliary_estimates=None):

Args:

• dataset (pd.DataFrame): Dataset with columns:
  • obs_init (int or NDArray[float]): Initial observation (state) $s_0$.
  • obs (int or NDArray[float]): Current observation (state) $s$.
  • act (int): Action $a$.
  • rew (float): Reward $R(s, a, s')$.
  • obs_next (int or NDArray[float]): Next observation (state) $s'$.
  • probs_next_evaluation or probs_next (NDArray[float]): Action probabilities under the target policy at the next state $\pi(\cdot \mid s')$.
• n_obs (int): Number of states $|S|$.
• n_act (int): Number of actions $|A|$.
• path (str, optional): Path to load/save estimates.
• verbosity (int): Verbosity level for logging.
• auxiliary_estimates (AuxiliaryEstimates, optional): Precomputed estimates.

If the auxiliary_estimates are not provided, a new instance of AuxiliaryEstimates will be instanciated.

from dice_rl_TU_Vienna.estimators.tabular.model_based import ModelBasedTabularOffPE

mb_solver = ModelBasedTabularOffPE(dataset, n_obs=20, n_act=5, path="./cache", verbosity=1)

# Access precomputed estimates
aux = mb_solver.auxiliary_estimates
d0_bar, dD_bar, P_bar, r_bar, n = aux.bar
d0_hat, dD_hat, P_hat, r_hat = aux.hat