def thompson_sample(mxmdl: omdict,
prior_a: float = 1,
prior_b: float = 1) -> tuple[Atom, float]:
"""Perform Thompson sampling over the mixture model.
Meaning, for each action
1. Select a TV according to its likelihood (derived from its
cognitive schematic).
2. From that TV, sample its second order distribution to obtain a
first order probability variate, and return the pair (action,
pblty) corresponding to the highest variate.
Then return the action with the highest probability of success.
"""
agent_log.fine("thompson_sample(mxmdl={}, prior_a={}, prior_b={})".format(
mxmdl_to_str(mxmdl), prior_a, prior_b))
# 1. For each action select its TV according its weight
act_w8d_cogscms = [(action, weighted_sampling(w8d_cogscms))
for (action, w8d_cogscms) in mxmdl.listitems()]
agent_log.fine("act_w8d_cogscms:\n{}".format(
act_w8d_cogscms_to_str(act_w8d_cogscms)))
# 2. For each action select its first order probability given its tv
act_pblts = [(action, tv_rv(get_cogscm_tv(w8_cogscm[1]), prior_a, prior_b))
for (action, w8_cogscm) in act_w8d_cogscms]
agent_log.fine("act_pblts:\n{}".format(act_pblts_to_str(act_pblts)))
# Return an action with highest probability of success (TODO: take
# case of ties)
return max(act_pblts, key=lambda act_pblt: act_pblt[1])
def mxmdl_to_str(mxmdl: omdict, indent: str = "") -> str:
"""Pretty print the given mixture model of cogscms"""
s = ""
for act_w8d_cogscms in mxmdl.listitems():
action = act_w8d_cogscms[0]
w8d_cogscms = act_w8d_cogscms[1]
s += "\n" + indent + str(action_to_str(action)) + "\n"
s += w8d_cogscms_to_str(w8d_cogscms, indent + " ")
return s
