tmaze

Contents

tmaze#

hybrid_aif.reference.envs.tmaze(
reward_probability=1.0,
punishment_probability=1.0,
cue_validity=0.95,
reward_condition=None,
)[source]#

The classic TMaze as a positional process and its parameters, tensor-identical to pymdp’s.

Two state factors (location, five levels, controlled; context, two levels, fixed) and three modalities (location, reward, cue; see the module docstring). reward_probability is the chance the rewarded arm pays, punishment_probability the chance the other arm punishes, cue_validity the chance the cue location shows the true context, and reward_condition fixes the context (None: 50/50). The probabilities may be traced, so batches of parameter sets build under vmap.