Be more explicit in README about DiscretePolicyEstimator

README.md

@@ -190,7 +190,7 @@ In most cases, one needs to sample complete trajectories. From a batch of trajec
 
 Training GFlowNets requires one or multiple estimators, called `GFNModule`s, which is an abstract subclass of `torch.nn.Module`. In addition to the usual `forward` function, `GFNModule`s need to implement a `required_output_dim` attribute, to ensure that the outputs have the required dimension for the task at hand; and some (but not all) need to implement a `to_probability_distribution` function.
 
-- `DiscretePolicyEstimator` is a `GFNModule` that defines the policies $P_F(. \mid s)$ and $P_B(. \mid s)$ for discrete environments. When `is_backward=False`, the required output dimension is `n = env.n_actions`, and when `is_backward=True`, it is `n = env.n_actions - 1`. These `n` numbers represent the logits of a Categorical distribution. The corresponding `to_probability_distribution` function transforms the logits by masking illegal actions (according to the forward or backward masks), then return a Categorical distribution. The masking is done by setting the corresponding logit to $-\infty$. The function also includes exploration parameters, in order to define a tempered version of $P_F$, or a mixture of $P_F$ with a uniform distribution.
+- `DiscretePolicyEstimator` is a `GFNModule` that defines the policies $P_F(. \mid s)$ and $P_B(. \mid s)$ for discrete environments. When `is_backward=False`, the required output dimension is `n = env.n_actions`, and when `is_backward=True`, it is `n = env.n_actions - 1`. These `n` numbers represent the logits of a Categorical distribution. The corresponding `to_probability_distribution` function transforms the logits by masking illegal actions (according to the forward or backward masks), then return a Categorical distribution. The masking is done by setting the corresponding logit to $-\infty$. The function also includes exploration parameters, in order to define a tempered version of $P_F$, or a mixture of $P_F$ with a uniform distribution. `DiscretePolicyEstimator`` with `is_backward=False`` can be used to represent log-edge-flow estimators $\log F(s \rightarrow s')$.
 - `ScalarModule` is a simple module with required output dimension 1. It is useful to define log-state flows $\log F(s)$.
 
 For non-discrete environments, the user needs to specify their own policies $P_F$ and $P_B$. The module, taking as input a batch of states (as a `States`) object, should return the batched parameters of a `torch.Distribution`. The distribution depends on the environment. The `to_probability_distribution` function handles the conversion of the parameter outputs to an actual batched `Distribution` object, that implements at least the `sample` and `log_prob` functions. An example is provided [here](https://github.com/saleml/torchgfn/tree/master/src/gfn/gym/helpers/box_utils.py), for a square environment in which the forward policy has support either on a quarter disk, or on an arc-circle, such that the angle, and the radius (for the quarter disk part) are scaled samples from a mixture of Beta distributions. The provided example shows an intricate scenario, and it is not expected that user defined environment need this much level of details.