infer-actively/pymdp
Design sensible defaults for sophisticated inference planning parameters
Offen
#342 geöffnet am 22.01.2026
enhancementhelp wanted
Repository-Metriken
- Stars
- (720 Sterne)
- PR-Merge-Metriken
- (PR-Metriken ausstehend)
Beschreibung
Follow up issue from discussion in review of #341
Right now in sophisticated inference (see si_policy_search in pymdp/planning/si.py), the default parameters are as follows:
def si_policy_search(
horizon=5,
max_nodes=5000,
max_branching=10,
policy_prune_threshold=1 / 16,
observation_prune_threshold=1 / 16,
entropy_stop_threshold=0.5,
efe_stop_threshold=1e10,
kl_threshold=-1,
prune_penalty=512,
gamma=1,
topk_obsspace=10000,
):
My intuition is that these parameters are incredibly model specific, and should depend on a combination of the A/B matrix parameters (which parameterizes the expected uncertainty associated with both observation and hidden state distributions), the planning horizon, and action space, to set them. One idea for this is to actual compute defaults from a given model, rather than set them as fixed numbers manually.