Antibias ignores the per-condition probability prior

[alexevag]

Antibias ignores the per-condition probability prior

Summary

When trial_selection is staircase or biased with antibias enabled, the realized condition distribution does not follow the experimenter-defined prior. With random the same prior is honored. Over 150 trials, staircase+antibias drifts toward a balanced/uniform distribution instead of the intended one.

Where

src/ethopy/core/experiment.py

• _random_selection (line ~680)
• _anti_bias (line ~635)
• _staircase_selection (line ~713)
• _biased_selection (line ~739)
• un_choices / un_blocks setup in log_conditions (line ~478)

Root cause

random honors the prior because it samples self.conditions directly:

return np.random.choice(self.conditions)

This is a uniform draw over the conditions list, whose multiplicity encodes the prior (A×3, B×3, C×4 → 30/30/40).

The antibias path discards that multiplicity in two places:

1. self.un_choices = np.unique(...) — antibias operates only on unique [response_port, difficulty] choices, so it can't see that C appeared 4× and A only 3×.
2. _anti_bias returns a choice with fixed_p = 1 - mean(recent == choice),normalized. That deliberately drives presentation toward equalizing the animal's recent choices (uniform target), not toward the experimenter prior.
3. The final np.random.choice(valid_conditions) is also uniform among the conditions matching the chosen choice.

result: with antibias the distribution depends on how many unique choices A/B/C map to and on the animal's behavior

This is "working as designed" for antibias (bias correction targets uniform), but it is incompatible with also reproducing a fixed prior.