NeuralGenDice estimates the policy value $\rho^\pi$ by first approximating the stationary distribution correction $w_{\pi/D}: S \times A \to R_{\geq 0}$ in the tabular setting. It is a gradient-based estimator that parameterizes the function arguments of the dual GenDICE objective with neural networks and performs stochastic gradient descent-ascent based on logged data. It inherits from NeuralDice, but overrides all the necessary base methods.

Unlike NeuralDualDice, NeuralGenDice supports both the discounted and undiscounted case, i.e., $0 < \gamma \leq 1$. However, for the latter, one must chose a positive norm regularization coefficient $\lambda > 0$.

Fenchel-Rockefeller duality is applied to the primal GenDICE objective:

$$ J(w) \doteq D_\phi( D^D w ~ | ~ \mathcal B^\pi_T D^D w ) + \frac{\lambda}{2} \left ( E_{ (s, a) \sim d^D } [ w(s, a) ] - 1 \right )^2, \quad \phi(x) = (x - 1)^2. $$

Note that:

• $D_\phi$ is the f-divergence based on $\phi$,
• $D^D$ is the diagonal matrix of $d^D$, the dataset distribution,
• $B^\pi_T$ is the backward Bellman operator,
• $\lambda$ is a Lagrange multiplier that penalizes deviations of the expectation of the stationary distribution candidate $w$ under the dataset distribution $d^D$ from $1$.

This yields the dual GenDICE objective:

$$ J(v, w) \doteq E_{ s_0 \sim d_0, ~ (s, a) \sim d^D, ~ s' \sim T(s, a) } \left[ L(v, w; s_0, s, a, s') + N(w, u; s, a) - \frac{1}{4} v(s, a)^2 w(s, a) \right ]. $$

The loss term $L(v, w; s_0, s, a, s')$ is the same as in DualDICE:

$$ L(v, w; s_0, s, a, s') \doteq (1 - \gamma) E_{ a_0 \sim \pi(s_0) } [ v(s_0, a_0) ] + w(s, a) \left ( \gamma E_{a' \sim \pi(s')} [ v(s', a') ] - v(s, a) \right ). $$

The new norm regularization term $N(w, u; s, a)$ and coefficient $\lambda$ for the divergence are:

$$ N(w, u; s, a) \doteq \lambda \left ( u ( w(s, a) - 1 ) - \frac{1}{2} u^2 \right ), \quad \lambda \geq 0. $$

The norm variable $u$ is an additional parameter learned by the model.

The optimization problem alternates between maximizing $J(v, w)$ with respect to $v$ and minimizing it with respect to $w$, along with the regularization and penalty terms. The stationary distribution correction $w_{\pi / D}$ is estimated based on the optimized $v^\ast$.

For further details, refer to

• the Bellman Equations wiki page,
• the original paper: GenDICE: Generalized Behavior-Agnostic Estimation of Discounted Stationary Distribution Corrections

def __init__(
        self,
        gamma, lamda, seed, batch_size,
        learning_rate, hidden_dimensions,
        obs_min, obs_max, n_act, obs_shape,
        dataset, preprocess_obs=None, preprocess_act=None, preprocess_rew=None,
        dir=None, get_recordings=None, other_hyperparameters=None, save_interval=100):

Args:

• All the arguments of NeuralDice are inherited.
• lamda (float): Norm regularization coefficient $\lambda$.

$\lambda$ is passed as lamda in the Python code due to the reserved keyword lambda in Python.

def get_loss(self, v_init, v, v_next, w):

Overrides the base class get_loss to compute the dual GenDICE objective.

def set_up_networks(self):

In addition to the base method, a tf.Variable and a corresponding SGD optimizer are set up for the norm variable $u$.

from some_module import NeuralGenDice

estimator = NeuralGenDice(
    gamma=0.99,
    lamda=0.5,
    seed=0,
    batch_size=64,
    learning_rate=1e-3,
    hidden_dimensions=(64, 64),
    obs_min=obs_min,
    obs_max=obs_max,
    n_act=4,
    obs_shape=(8,),
    dataset=df,
    dir="./logs"
)

estimator.evaluate_loop(n_steps=10_000)

rho_hat = estimator.solve_pv(weighted=True)