loss = 1 - agg_p(truth_values) – the standard LTN "Real Logic" training objective (research_2026_neuro_symbolic_ai.md §1/§2): maximizing a knowledge base's aggregated satisfaction via ordinary gradient descent is the same as minimizing this loss. More...

#include <satisfaction_loss.hpp>

Public Member Functions
	SatisfactionLoss (DeviceBackend *backend, float p=2.0f)
	Constructs a satisfaction loss with an internally-owned p-mean aggregator.

float	forward (const Tensor &truth_values)
	Aggregates truth_values into a single scalar satisfaction degree and returns `1 - that`.

Tensor	backward ()
	Gradient w.r.t. truth_values: `d(loss)/d(sat) == -1` exactly (the complement is affine), propagated back through the internally-owned AggregatorModule's own backward().

Detailed Description

loss = 1 - agg_p(truth_values) – the standard LTN "Real Logic" training objective (research_2026_neuro_symbolic_ai.md §1/§2): maximizing a knowledge base's aggregated satisfaction via ordinary gradient descent is the same as minimizing this loss.

Note: Design decision (mission_1_aggregator_satisfaction_loss.md Stage 3, resolved): not a Module subclass – confirmed directly against mse_loss.hpp's own stated rationale, not assumed by analogy alone. MSELoss's own header states losses are "the seed point relevance/gradient propagation starts from, not something a propagate_relevance rule is defined for – LRP explains a model's prediction, not the loss function used to train it." That rationale is about role (a loss sits outside the relevance-bearing forward computation entirely, by definition), not about argument count – so it applies identically here even though this loss's arity differs from MSELoss's. MSELoss takes exactly two tensors (prediction, target); SatisfactionLoss takes exactly one (a single formula's already-composed per-grounding truth degrees) – a genuinely different input shape, checked here rather than silently assumed identical, but the same "loss, not Module" classification either way.; Composes Mission 0's ConjunctionModule/DisjunctionModule/NegationModule with this mission's AggregatorModule the way every multi-module pipeline in this codebase composes – by chaining calls at the call site, not by one class owning every piece (e.g. XorNetwork chains LinearModule/ReluModule externally rather than a monolithic class inlining both). This class's own single responsibility is the aggregation-and-complement step (1 - agg_p(...)) – it owns exactly one AggregatorModule instance. Grounding a specific first-order formula by chaining Conjunction/Disjunction/Negation calls over predicate outputs happens upstream, at the call site; which concrete formula that is is Phase 1 Mission 2's own scope (the toy knowledge base), not re-implemented or anticipated here.; Scope restriction: truth_values must be rank 1 – a single formula's per-grounding truth degrees, batch axis only, no other dimensions. A multi-formula knowledge base (several formulas' satisfaction degrees combined into one training signal) is Mission 2's scope, not this class's; documented explicitly here rather than silently generalized ahead of that mission's own design work.

Constructor & Destructor Documentation

◆ SatisfactionLoss()

pulsatrix::SatisfactionLoss::SatisfactionLoss	(	DeviceBackend *	backend,
		float	p = `2.0f`
	)

explicit

Constructs a satisfaction loss with an internally-owned p-mean aggregator.

Parameters

backend	Backend to allocate/compute through. Not owned; must outlive this loss.
p	Aggregator exponent, forwarded to AggregatorModule (see its own constructor note on p == 0 being rejected there). Defaults to 2.0, matching AggregatorModule's own default.

Member Function Documentation

◆ backward()

Tensor pulsatrix::SatisfactionLoss::backward ( )

Gradient w.r.t. truth_values: d(loss)/d(sat) == -1 exactly (the complement is affine), propagated back through the internally-owned AggregatorModule's own backward().

Returns: Gradient tensor, same shape as the truth_values passed to forward().

Exceptions

std::logic_error if called before any forward().

Note: Not const – unlike MSELoss::backward(), which only reads cached members, this method delegates to AggregatorModule::backward(), which Module declares non-const (it is a virtual method every mutable-state subclass overrides). Documented explicitly as a deliberate deviation from MSELoss's own const signature, not an oversight.

◆ forward()

float pulsatrix::SatisfactionLoss::forward ( const Tensor & truth_values )

Aggregates truth_values into a single scalar satisfaction degree and returns 1 - that.

Parameters

truth_values Rank-1 tensor of per-grounding formula-truth degrees. Must be rank 1 – see the class note on scope.

Returns: 1 - agg_p(truth_values), the scalar loss to minimize via gradient descent.