Docs / Pragmatics / Probabilistic / Decision

The decision-theoretic listener and speaker

import Linglib.Pragmatics.Probabilistic.Decision · namespace RSA

A listener who acts holds a belief after an utterance, and a decision problem gives a reward for each action at each state. Her policy is the softmax of an action's expected reward under her belief, and the decision-theoretic speaker the softmax of the true reward of the listener's choice, in expectation, less a cost. The rule is Kernel.softmax, as for RSA.speaker, and only the utility differs: hers is linear in the listener's choice where RSA.utility is the logarithm of the listener's mass, so she sends every utterance, false ones included.

References

Declarations

The listener's policy is the softmax at rationality β of an action's expected reward under her belief.

policyL40–43
noncomputable def RSA.policy {W : Type u_1} {U : Type u_2} {A : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [Fintype U] [DiscreteMeasurableSpace U] [Fintype A] [MeasurableSpace A] (β : ℝ) (L : ProbabilityTheory.Kernel U W) (R : A → W → ℝ) :
ProbabilityTheory.Kernel U A

Used by instIsFiniteKernelPolicy, instIsMarkovKernelPolicyOfNonempty, policy_real_singleton, policy_real_lt_iff, policy_single_real_lt_iff, policy_real_of_pair, integral_policy_eq_of_integral_eq_add

Used in HardingGerstenbergIcard2025, QingFranke2015, SumersEtAl2024

L47
L49–51
L55–58
L60–65
L67–74
L76–83
L85–92

The decision-theoretic speaker

The decision-theoretic speaker is the softmax at rationality β of the true reward of the listener's choice, in expectation, less the utterance's cost. Her utility is linear in the listener's probabilities, where RSA.utility is the logarithm of the listener's mass.

actionSpeakerL103–108
noncomputable def RSA.actionSpeaker {W : Type u_1} {M : Type u_2} {A : Type u_3} [MeasurableSpace W] [Countable W] [MeasurableSingletonClass W] [MeasurableSpace M] [Fintype M] [MeasurableSpace A] (β : ℝ) (cost : M → ℝ) (π : ProbabilityTheory.Kernel M A) (R : A → W → ℝ) :
ProbabilityTheory.Kernel W M

Used by instIsFiniteKernelActionSpeaker, instIsMarkovKernelActionSpeakerOfNonempty, actionSpeaker_apply_singleton_ne_zero, actionSpeaker_real_lt_iff, actionSpeaker_real_of_pair

Used in BrochhagenFrankeVanRooij2018, HardingGerstenbergIcard2025

L112–113
L115–117

The speaker sends every utterance, false ones included, with positive probability.

theorem actionSpeaker_apply_singleton_ne_zeroL121–123
{W : Type u_1} {M : Type u_2} {A : Type u_3} [MeasurableSpace W] [Countable W] [MeasurableSingletonClass W] [MeasurableSpace M] [Fintype M] [MeasurableSpace A] {β : ℝ} (R : A → W → ℝ) (cost : M → ℝ) (π : ProbabilityTheory.Kernel M A) {m : M} {w : W} [MeasurableSingletonClass M] :
((actionSpeaker β cost π R) w) {m} ≠ 0

Built from

L125–129
L131–139