wafrift_evolution/search/mod.rs
1//! Search algorithms for evolutionary WAF bypass discovery.
2
3use crate::evolution::{Chromosome, GenePool};
4use crate::types::{Budget, EvolutionError, OracleVerdict, SearchStats};
5use rand::rngs::StdRng;
6use std::cmp::Ordering;
7
8/// A candidate requested for evaluation, with a stable evaluation ID.
9#[derive(Debug, Clone)]
10pub struct EvalCandidate {
11 /// Stable ID used to correlate results.
12 pub id: u64,
13 /// The chromosome to evaluate.
14 pub chromosome: Chromosome,
15}
16
17/// Result of submitting evaluations back to the algorithm.
18#[derive(Debug, Clone, Copy, PartialEq, Eq)]
19pub enum SubmitResult {
20 /// The algorithm accepted all results.
21 Accepted,
22 /// Some evaluation IDs were unknown.
23 UnknownIds(usize),
24}
25
26/// Core trait implemented by all search algorithms.
27///
28/// Each algorithm manages its own internal state (population, archive,
29/// temperature, tabu list, etc.). The [`EvolutionEngine`](crate::evolution::EvolutionEngine)
30/// handles caching, budgeting, and batching on top of this trait.
31pub trait SearchAlgorithm: Send + Sync + std::fmt::Debug {
32 /// Algorithm name.
33 fn name(&self) -> &'static str;
34
35 /// Initialize the algorithm with a seed population.
36 fn initialize(&mut self, population: Vec<Chromosome>, gene_pool: &GenePool, rng: &mut StdRng);
37
38 /// Request up to `n` candidates for parallel evaluation.
39 ///
40 /// Returns candidates with stable IDs. The caller evaluates them and
41 /// later calls [`submit_evaluations`](SearchAlgorithm::submit_evaluations).
42 fn request_evaluations(&mut self, n: usize, rng: &mut StdRng) -> Vec<EvalCandidate>;
43
44 /// Submit evaluation results.
45 ///
46 /// The ID in each tuple must match an ID previously returned by
47 /// `request_evaluations`.
48 fn submit_evaluations(&mut self, results: Vec<(u64, OracleVerdict)>);
49
50 /// Check whether the algorithm thinks search should stop.
51 fn should_terminate(&self, stats: &SearchStats, budget: &Budget) -> bool;
52
53 /// Get the best chromosome found so far.
54 fn best(&self) -> Option<&Chromosome>;
55
56 /// Serialize internal state to bytes.
57 fn checkpoint(&self) -> Result<Vec<u8>, EvolutionError>;
58
59 /// Restore internal state from bytes.
60 fn restore(&mut self, bytes: &[u8]) -> Result<(), EvolutionError>;
61
62 /// Snapshot the algorithm's current "live" chromosomes — the set
63 /// the engine is actively searching from.
64 ///
65 /// Population-based algorithms (`NoveltySearch`, `MapElites`)
66 /// return their full pool. Single-state algorithms (`HillClimbing`,
67 /// `SimulatedAnnealing`, `TabuSearch`) return the singleton
68 /// current/best so the engine sees `len() == 1` and reports
69 /// "no pairwise diversity yet" rather than zero.
70 ///
71 /// Used by [`EvolutionEngine::diversity_score`](crate::evolution::EvolutionEngine::diversity_score)
72 /// to drive adaptive mutation pressure. Cloning is acceptable here
73 /// because callers run it at engine-tick rate, not per-evaluation.
74 fn population_snapshot(&self) -> Vec<Chromosome> {
75 self.best().cloned().into_iter().collect()
76 }
77
78 /// Deep-clone this algorithm into a fresh trait object.
79 ///
80 /// The default implementation falls back to a `checkpoint` →
81 /// `restore` round-trip via serde_json — correct for any
82 /// algorithm that implements those, but slow on large grids /
83 /// archives because every chromosome is JSON-serialised.
84 ///
85 /// Concrete algorithms override this with a direct in-memory
86 /// `Clone` to bypass the serde round-trip — typically 10-100×
87 /// faster on populated state. The override is what
88 /// [`EvolutionEngine::clone`](crate::evolution::EvolutionEngine)
89 /// uses on the proxy path, where allocation spikes from JSON
90 /// were the original blocker.
91 fn clone_box(&self) -> Box<dyn SearchAlgorithm> {
92 // The trait can't construct a fresh same-typed instance
93 // without help from the algorithm registry, so the default
94 // implementation is intentionally a panic for out-of-tree
95 // algorithms — they should override `clone_box` with a real
96 // `Clone`-based path. The 5 in-tree algorithms always
97 // override.
98 panic!(
99 "default clone_box is unreachable for in-tree algorithms; \
100 out-of-tree algorithms must override this method"
101 );
102 }
103}
104
105/// Convert non-finite fitness values into a strict worst-case sentinel.
106///
107/// NaN and +/-inf break partial ordering semantics and can lock algorithms
108/// into never-accept states. Mapping them to `-inf` keeps comparisons total.
109#[must_use]
110pub(crate) fn comparable_fitness(value: f64) -> f64 {
111 if value.is_finite() {
112 value
113 } else {
114 f64::NEG_INFINITY
115 }
116}
117
118#[must_use]
119pub(crate) fn fitness_cmp(a: f64, b: f64) -> Ordering {
120 comparable_fitness(a)
121 .partial_cmp(&comparable_fitness(b))
122 .unwrap_or(Ordering::Equal)
123}
124
125pub mod hill_climb;
126pub mod map_elites;
127pub mod novelty;
128pub mod sim_anneal;
129pub mod tabu;
130
131pub use hill_climb::HillClimbing;
132pub use map_elites::MapElites;
133pub use novelty::NoveltySearch;
134pub use sim_anneal::SimulatedAnnealing;
135pub use tabu::TabuSearch;