Owl_rllib.NormaliseObservation and reward normalisation, as Gymnasium 0.29.1's NormalizeObservation and NormalizeReward.
Statistics are float64 running moments (starting from mean 0, variance 1, count 1e-4), updated with every value seen. Each wrapped environment keeps its own statistics.
val reward :
?epsilon:float ->
?bound:float ->
gamma:float ->
('obs, 'action) Ndenv.t ->
('obs, 'action) Ndenv.tDivide rewards by the running standard deviation of the discounted return, then clip them to [-bound, bound] (default: no clip). As in Gymnasium, the return restarts after a termination but not after a truncation. epsilon defaults to 1e-8.
module S : sig ... endmodule D : sig ... end