Skip to main content

elevator_core/dispatch/
etd.rs

1//! Estimated Time to Destination (ETD) dispatch algorithm.
2//!
3//! The per-call cost-minimization approach is drawn from Barney, G. C. &
4//! dos Santos, S. M., *Elevator Traffic Analysis, Design and Control* (2nd
5//! ed., 1985). Commercial controllers (Otis Elevonic, KONE Polaris, etc.)
6//! use variants of the same idea; this implementation is a simplified
7//! educational model, not a faithful reproduction of any vendor's system.
8
9use smallvec::SmallVec;
10
11use crate::components::{ElevatorPhase, Route};
12use crate::entity::EntityId;
13use crate::world::World;
14
15use super::{DispatchManifest, DispatchStrategy, ElevatorGroup, RankContext, pair_can_do_work};
16
17/// Estimated Time to Destination (ETD) dispatch algorithm.
18///
19/// For each `(car, stop)` pair the rank is a cost estimate combining
20/// travel time, delay imposed on riders already aboard, door-overhead
21/// for intervening stops, and a small bonus for cars already heading
22/// toward the stop. The dispatch system runs an optimal assignment
23/// across all pairs so the globally best matching is chosen.
24pub struct EtdDispatch {
25    /// Weight for travel time to reach the calling stop.
26    pub wait_weight: f64,
27    /// Weight for delay imposed on existing riders.
28    pub delay_weight: f64,
29    /// Weight for door open/close overhead at intermediate stops.
30    pub door_weight: f64,
31    /// Weight for the squared-wait "group-time" fairness bonus. Each
32    /// candidate stop's cost is reduced by this weight times the sum
33    /// of `wait_ticks²` across waiting riders at the stop, so stops
34    /// hosting older calls win ties. Defaults to `0.0` (no bias);
35    /// positive values damp the long-wait tail (Aalto EJOR 2016
36    /// group-time assignment model).
37    pub wait_squared_weight: f64,
38    /// Weight for the linear waiting-age fairness term. Each candidate
39    /// stop's cost is reduced by this weight times the sum of
40    /// `wait_ticks` across waiting riders at the stop, so stops hosting
41    /// older calls win ties without the quadratic blow-up of
42    /// [`wait_squared_weight`](Self::wait_squared_weight). Defaults to
43    /// `0.0` (no bias); positive values implement the linear
44    /// collective-group-control fairness term from Lim 1983 /
45    /// Barney–dos Santos 1985 CGC.
46    ///
47    /// Composes additively with `wait_squared_weight`: users wanting
48    /// the full CGC shape can set both (`k·Σw + λ·Σw²`).
49    pub age_linear_weight: f64,
50    /// Positions of every demanded stop in the group, cached by
51    /// [`DispatchStrategy::pre_dispatch`] so `rank` avoids rebuilding the
52    /// list for every `(car, stop)` pair.
53    pending_positions: SmallVec<[f64; 16]>,
54}
55
56impl EtdDispatch {
57    /// Create a new `EtdDispatch` with default weights.
58    ///
59    /// Defaults: `wait_weight = 1.0`, `delay_weight = 1.0`,
60    /// `door_weight = 0.5`, `wait_squared_weight = 0.0`,
61    /// `age_linear_weight = 0.0`.
62    #[must_use]
63    pub fn new() -> Self {
64        Self {
65            wait_weight: 1.0,
66            delay_weight: 1.0,
67            door_weight: 0.5,
68            wait_squared_weight: 0.0,
69            age_linear_weight: 0.0,
70            pending_positions: SmallVec::new(),
71        }
72    }
73
74    /// Create with a single delay weight (backwards-compatible shorthand).
75    #[must_use]
76    pub fn with_delay_weight(delay_weight: f64) -> Self {
77        Self {
78            wait_weight: 1.0,
79            delay_weight,
80            door_weight: 0.5,
81            wait_squared_weight: 0.0,
82            age_linear_weight: 0.0,
83            pending_positions: SmallVec::new(),
84        }
85    }
86
87    /// Create with fully custom weights.
88    #[must_use]
89    pub fn with_weights(wait_weight: f64, delay_weight: f64, door_weight: f64) -> Self {
90        Self {
91            wait_weight,
92            delay_weight,
93            door_weight,
94            wait_squared_weight: 0.0,
95            age_linear_weight: 0.0,
96            pending_positions: SmallVec::new(),
97        }
98    }
99
100    /// Turn on the squared-wait fairness bonus. Higher values prefer
101    /// older waiters more aggressively; `0.0` (the default) disables.
102    ///
103    /// # Panics
104    /// Panics on non-finite or negative weights. A `NaN` weight would
105    /// propagate through `mul_add` and silently disable every dispatch
106    /// rank; a negative weight would invert the fairness ordering.
107    /// Either is a programming error rather than a valid configuration.
108    #[must_use]
109    pub fn with_wait_squared_weight(mut self, weight: f64) -> Self {
110        assert!(
111            weight.is_finite() && weight >= 0.0,
112            "wait_squared_weight must be finite and non-negative, got {weight}"
113        );
114        self.wait_squared_weight = weight;
115        self
116    }
117
118    /// Turn on the linear waiting-age fairness term. Higher values
119    /// prefer older waiters more aggressively; `0.0` (the default)
120    /// disables. Composes additively with
121    /// [`with_wait_squared_weight`](Self::with_wait_squared_weight).
122    ///
123    /// # Panics
124    /// Panics on non-finite or negative weights, for the same reasons
125    /// as [`with_wait_squared_weight`](Self::with_wait_squared_weight).
126    #[must_use]
127    pub fn with_age_linear_weight(mut self, weight: f64) -> Self {
128        assert!(
129            weight.is_finite() && weight >= 0.0,
130            "age_linear_weight must be finite and non-negative, got {weight}"
131        );
132        self.age_linear_weight = weight;
133        self
134    }
135}
136
137impl Default for EtdDispatch {
138    fn default() -> Self {
139        Self::new()
140    }
141}
142
143impl DispatchStrategy for EtdDispatch {
144    fn pre_dispatch(
145        &mut self,
146        group: &ElevatorGroup,
147        manifest: &DispatchManifest,
148        world: &mut World,
149    ) {
150        self.pending_positions.clear();
151        for &s in group.stop_entities() {
152            if manifest.has_demand(s)
153                && let Some(p) = world.stop_position(s)
154            {
155                self.pending_positions.push(p);
156            }
157        }
158    }
159
160    fn rank(&mut self, ctx: &RankContext<'_>) -> Option<f64> {
161        // Exclude `(car, stop)` pairs that can't produce any useful work.
162        // Without this guard, a full car whose only candidate stop is a
163        // pickup it lacks capacity to serve collapses to a zero-cost
164        // self-assignment (travel, detour, and door terms are all 0 when
165        // the car is already at the stop). Dispatch then re-selects that
166        // stop every tick — doors cycle open, reject, close, repeat — and
167        // the aboard riders are never carried to their destinations.
168        if !pair_can_do_work(ctx) {
169            return None;
170        }
171        let mut cost = self.compute_cost(ctx.car, ctx.car_position, ctx.stop_position, ctx.world);
172        if self.wait_squared_weight > 0.0 {
173            let wait_sq: f64 = ctx
174                .manifest
175                .waiting_riders_at(ctx.stop)
176                .iter()
177                .map(|r| {
178                    let w = r.wait_ticks as f64;
179                    w * w
180                })
181                .sum();
182            cost = self.wait_squared_weight.mul_add(-wait_sq, cost).max(0.0);
183        }
184        if self.age_linear_weight > 0.0 {
185            let wait_sum: f64 = ctx
186                .manifest
187                .waiting_riders_at(ctx.stop)
188                .iter()
189                .map(|r| r.wait_ticks as f64)
190                .sum();
191            cost = self.age_linear_weight.mul_add(-wait_sum, cost).max(0.0);
192        }
193        if cost.is_finite() { Some(cost) } else { None }
194    }
195}
196
197impl EtdDispatch {
198    /// Compute ETD cost for assigning an elevator to serve a stop.
199    ///
200    /// Cost = `wait_weight` * travel\_time + `delay_weight` * existing\_rider\_delay
201    ///      + `door_weight` * door\_overhead + direction\_bonus
202    fn compute_cost(
203        &self,
204        elev_eid: EntityId,
205        elev_pos: f64,
206        target_pos: f64,
207        world: &World,
208    ) -> f64 {
209        let Some(car) = world.elevator(elev_eid) else {
210            return f64::INFINITY;
211        };
212
213        let distance = (elev_pos - target_pos).abs();
214        let travel_time = if car.max_speed.value() > 0.0 {
215            distance / car.max_speed.value()
216        } else {
217            return f64::INFINITY;
218        };
219
220        // Door overhead is a seconds-denominated cost so the Hungarian
221        // can compare it apples-to-apples against travel time and
222        // existing-rider delay. Pre-fix, this was summed in ticks,
223        // multiplied by `door_weight` (dimensionless), and added to
224        // seconds-valued terms — giving door cost ~60× the intended
225        // influence at 60 Hz. A single intervening stop could then
226        // outweigh a long travel time and bias ETD toward distant
227        // cars with clear shafts over closer ones with a single
228        // waypoint. Convert with the sim's tick rate (resource-
229        // provided) and fall back to 60 Hz for bare-World contexts
230        // such as unit-test fixtures.
231        let tick_rate = world
232            .resource::<crate::time::TickRate>()
233            .map_or(60.0, |r| r.0);
234        let door_overhead_per_stop =
235            f64::from(car.door_transition_ticks * 2 + car.door_open_ticks) / tick_rate;
236
237        // Intervening pending stops between car and target contribute door overhead.
238        let (lo, hi) = if elev_pos < target_pos {
239            (elev_pos, target_pos)
240        } else {
241            (target_pos, elev_pos)
242        };
243        let intervening_stops = self
244            .pending_positions
245            .iter()
246            .filter(|p| **p > lo + 1e-9 && **p < hi - 1e-9)
247            .count() as f64;
248        let door_cost = intervening_stops * door_overhead_per_stop;
249
250        let mut existing_rider_delay = 0.0_f64;
251        for &rider_eid in car.riders() {
252            if let Some(dest) = world.route(rider_eid).and_then(Route::current_destination)
253                && let Some(dest_pos) = world.stop_position(dest)
254            {
255                let direct_dist = (elev_pos - dest_pos).abs();
256                let detour_dist = (elev_pos - target_pos).abs() + (target_pos - dest_pos).abs();
257                let extra = (detour_dist - direct_dist).max(0.0);
258                if car.max_speed.value() > 0.0 {
259                    existing_rider_delay += extra / car.max_speed.value();
260                }
261            }
262        }
263
264        // Direction bonus: if the car is already heading this way, subtract.
265        // Scoring model requires non-negative costs, so clamp at zero — losing
266        // a small amount of discriminative power vs. a pure free-for-all when
267        // two assignments tie.
268        let direction_bonus = match car.phase.moving_target() {
269            Some(current_target) => world.stop_position(current_target).map_or(0.0, |ctp| {
270                let moving_up = ctp > elev_pos;
271                let target_is_ahead = if moving_up {
272                    target_pos > elev_pos && target_pos <= ctp
273                } else {
274                    target_pos < elev_pos && target_pos >= ctp
275                };
276                if target_is_ahead {
277                    -travel_time * 0.5
278                } else {
279                    0.0
280                }
281            }),
282            None if car.phase == ElevatorPhase::Idle => -travel_time * 0.3,
283            _ => 0.0,
284        };
285
286        let raw = self.wait_weight.mul_add(
287            travel_time,
288            self.delay_weight.mul_add(
289                existing_rider_delay,
290                self.door_weight.mul_add(door_cost, direction_bonus),
291            ),
292        );
293        raw.max(0.0)
294    }
295}