Hotel Swarm Operations
Four everyday problems in a five-hotel group, each handed to a team of small cooperating decision-makers (a "swarm"), then measured against how hotels usually handle them today. Every result is from data or scenarios the swarm had not seen while it was being built.
Scorecard
Each problem compared three ways: what many hotels do today, the best simple rule a manager could write down, and the swarm. Green means the swarm did best; amber means it did not beat the best rule yet.
| Problem | Measured as | Today's practice | Best simple rule | Swarm | Verdict |
|---|
What a "swarm" means here
Instead of one central planner, every piece of the operation makes its own small decision: each chiller and laundry line, each room attendant, each displaced reservation, each employee. They share a common goal and a few simple signals, so together they act like one well-coordinated team.
This suits hotels because the decisions are many, small, local and time-sensitive. It also scales: the same setup works for one property or fifty.
Two of the four problems use a swarm that learned from practice runs (energy, housekeeping). The other two use a swarm that negotiates through bids, like a quick internal auction (relocation, shifts). Both kinds were checked against the best possible plan where one can be calculated.
How to read this dashboard
- Today's practice is the common status quo (do nothing special, clean floor by floor, walk guests to a competitor, call an agency).
- Best simple rule is a sensible hand-written policy a manager could put in an SOP. The swarm has to beat this to be worth it.
- Best possible plan (where shown) is a mathematical optimum computed with full hindsight. Real operations can't reach it, but it shows how much is left on the table.
- Dollar figures depend on the stated assumptions (rates, costs, equipment sizes). Each tab lists them, and they are meant to be swapped for the group's own numbers.
The five sample hotels
A fictional group used throughout this portfolio, sized like a mid-size destination market. The equipment counted as "flexible" is what can safely shift its electricity use by an hour or two.
| Hotel | Style | Rooms | Base rate | Flexible equipment | Load in peak hours (kW) | Of which can shift (kW) |
|---|
The problem. In Ontario, a large electricity customer on the "Class A" rate pays the Global Adjustment (often the biggest line on the bill) based on its share of the province's five highest-demand hours of the year. Use less power in just those five hours and the charge drops for the whole following year. Nobody knows in advance which five hours they will be, only that they come on hot summer weekday afternoons.
The swarm. Each chiller, laundry line, EV charger and pool heater in the five hotels decides every hour whether to run early (pre-cool, pre-heat, catch up), run normally, or ease off. It sees what an engineer could see: current and forecast grid demand, the weather, the time and its own state. Guest comfort and daily work are protected: laundry and EV charging must be finished by the end of the day, and cooling only coasts on what was stored earlier.
The hottest day: what the hotels drew from the grid, hour by hour
All five hotels combined, on the held-out day with the highest grid demand. The shaded hours are among the five that set next year's charge. Hover or tap for values.
Read it like this: where a line rises above "no action", equipment is running early (pre-cooling, laundry ahead of schedule). Where it dips below, equipment is easing off. The rule pre-cools in a fixed late-morning block. The swarm builds its cooling reserve gradually through the day, then holds the hotels lowest right through the shaded peak hours. Both catch up in the evening, once the grid peak has passed, and evening use does not affect the charge.
The five hours that set the bill
Combined hotel load in each of the season's five highest grid hours (kW). Lower is better.
Saving by hotel, per year
Global Adjustment saved, from each hotel's average cut across the five peak hours.
Which equipment does the work
Average load taken off the five peak hours, by type of equipment, all hotels combined (kW).
Chillers carry most of the cut because a building can be pre-cooled and then coast for a few hours without guests noticing. Laundry is the easiest to move: running it after 6 pm costs nothing.
A typical summer day: when load moves
Average change to the hotels' load by hour of day, across all held-out days (kW). Above zero = running earlier or catching up; below zero = easing off.
All approaches compared
Held-out summer days (every other week of June to September, never used for learning). Yearly figures for the five hotels together.
| Approach | Cut in the 5 peak hours | Share of hotel load | Peak hours hit | Global Adjustment saved | Extra energy cost | Net saving per year | Laundry / charging left undone |
|---|
Every weekday runs the pre-cool and ease-off routine every weekday. Peak-alert rule runs it only on days forecast to be in the top 20% (similar to the IESO Peak Tracker alerts). The swarm wins on the cut itself. It spends more on extra energy because it pre-cools harder, but still comes out about a year ahead of the best rule.
Eligibility check. Class A is for accounts with average peak demand above 1 MW; accounts between 500 kW and 1 MW can opt in. In this sample only Summit Tower and Grand Terrace are large enough to opt in on their own, which would be a year of the saving. The others would need a different tariff lever, such as a demand charge or a demand-response program, where the same swarm applies.
Assumptions
- Global Adjustment value of a kW cut in the five peaks: $450 per kW-year. Extra energy priced at $0.11/kWh.
- Equipment sizes are engineering estimates scaled by room count (chiller 0.3 kW/room, laundry per occupied room, one 7 kW EV charger per 60 rooms, pool heating by style). Replace them with each property's equipment list.
- Grid demand pattern: a public hourly summer demand record (2020) stands in for Ontario's IESO demand data. The IESO feed and Environment Canada weather plug straight in.
- Learned on half of the summer's weeks and tested on the other half.
The problem. Every morning about a third of the rooms check out, mostly between 9 and 11, and the time isn't known until it happens. Arriving guests want their room at check-in, and some paid for early check-in. Cleaning floor by floor is easy to manage but leaves guests in the lobby while rooms they need sit dirty on another floor.
The swarm. Each room attendant, when free, looks at the dirty rooms that matter most from where they stand (who arrives soonest, early check-in promises, the elevator trip) and picks the next room. Attendants never double up on a room. A sample day: a 300-room hotel on 10 floors, 8 attendants, 08:00 to 17:00, 30 minutes per room plus 10 to change floor, 15% of arriving guests on paid early check-in.
Guest waiting time per day
Total hours guests waited for a room after arriving, in one hotel per day (an early check-in guest's wait counts three times). Lower is better.
Early check-in promises kept
Share of paid early check-ins whose room was ready when the guest arrived. Higher is better.
All approaches compared
Averages over many simulated days with different check-out and arrival times.
| Approach | Guest waiting (hours/day) | Early check-ins on time | Rooms ready by noon | Workload evenness rooms per attendant, spread | Floor changes per day |
|---|
What this means for a hotel
The big win is the change of mindset, from "clean the floor" to "clean the room the next guest needs". That alone cuts guest waiting by about , and both the simple rule and the swarm get there.
Floor order still readies more rooms by noon because it never loses time on the elevator. But those are often rooms nobody is waiting for yet.
Honest result. The learned swarm comes close to the "soonest needed" rule but does not beat it yet: . For a single hotel, the rule is the practical choice today, and it can go into a housekeeping app or SOP right away.
The swarm earns its place when real-world friction is added: attendants on different sections and skills, rush requests, VIPs, rooms blocked for maintenance. A fixed rule can't weigh those trade-offs, but the swarm learns them. That is the next experiment.
The problem. On an oversold night, the front desk has to "walk" some guests. Usually that means paying for a competitor's room and transport, losing the revenue, and risking a loyal guest's future stays. With five hotels in the same market, many of those guests could stay inside the group instead, if someone quickly matches each reservation to the right sister hotel.
The swarm. Every reservation that has to move bids for the open rooms at the sister hotels, based on what the move is worth to the group (revenue kept, minus transport, a goodwill credit and any downgrade). The rooms go to the highest bids in a fast internal auction. The hard rules are always respected: a group stays together, an accessible-room guest only gets an accessible room, and a multi-night guest only goes where every night is available.
Result per oversold night
Net value to the group: revenue kept, minus every cost of moving or walking guests. Higher is better.
Where the money goes
Revenue kept in the group versus costs paid out, per oversold night.
All approaches compared
Averages over 50 simulated oversold nights. In each one, a random hotel in the group oversells by about 8%, and the sister hotels have about 2% of their rooms free.
| Approach | Net value per night | Revenue kept | Costs paid | Guests kept in the group | Average transfer | Rule breaks |
|---|
What this means for a hotel group
Keeping walked guests in the group turns a loss of about per oversold night into a gain. The auction adds another per night over "send them to the nearest sister hotel", because it puts each guest where they are worth the most. Long stays and groups go where every night is available, and short stays fill the gaps.
It reaches of the best possible plan, and it decides in a fraction of a second, so it can run at the front desk.
Assumptions
- Walking a guest costs the competitor's room (1.1× their rate), $40 transport and 20% of a loyal guest's yearly value.
- Moving inside the group costs $15 plus $2/km transport and a $50 goodwill credit. A suite guest moved to a regular room keeps 85% of the rate.
- 30% of displaced reservations are groups of 2 to 5 rooms, 5% of individual guests need an accessible room, and 40% stay more than one night.
- Hotel positions are rough distances (up to about 17 km apart) within one market.
The problem. Every week, shifts open up across the group: call-outs, busy nights, events. The usual fixes are calling an agency (expensive) or offering the shift to whoever answers first (fast, but it can create overtime and lean on the same few people).
The swarm. Each open shift is announced to staff across all five hotels. Every employee who is qualified, rested and under the weekly limit bids their true cost: their pay, overtime if it applies, travel to a sister hotel, plus a fairness charge for every extra shift they've already picked up. The lowest bid wins, and if no one beats the agency price, the shift goes to the agency.
Cost to cover a week's open shifts
25 open eight-hour shifts across the group. Lower is better.
Overtime and travel created
Overtime hours and shifts worked at a sister hotel, per week.
All approaches compared
Averages over 50 simulated weeks, each with a different roster and different open shifts.
| Approach | Weekly cost | Filled by own staff | Agency hours | Overtime hours | Cross-hotel shifts | Most extra shifts on one person |
|---|
What this means for a hotel group
Filling open shifts with the group's own people instead of an agency saves about a week. Compared with "first to accept", the market saves another a week (about a year) with less overtime and fewer cross-town trips, and it gets within of the best possible plan.
The trade-off is visible in the table: the market sends slightly more shifts to the agency when the only available staff would go into costly overtime.
Rules built in (Ontario Employment Standards Act)
- No bid if the shift would push someone past 48 hours in the week (s.17).
- At least 8 hours off between shifts, counting travel time (s.18).
- Overtime at time and a half after 44 hours (s.22).
- Only staff qualified for the department can bid (housekeeping, front desk, food and beverage).
Assumptions: wages $18 to $26/h, agency $40/h all-in, $25 plus 30 paid minutes for a cross-hotel shift, $20 fairness charge per extra shift already taken.
Weekly schedule: optimiser versus a fixed pattern
The other half of staffing: building the base schedule itself. This uses the optimiser behind the Staff Schedule Builder, which matches staff to forecast demand and honours time-off requests. Shortfall hours are hours with fewer people on than demand needs.
| Week | Method | Labour cost | Shortfall hours | Time-off requests honoured | Rule breaks |
|---|
From hotel data to a decision
- Describe the decisionWho decides (a chiller, an attendant, a reservation, an employee), what they can see, what they may do, and the hard rules.
- Set the goal in dollars or minutesPeak charges saved, guest minutes waited, revenue kept, labour cost. The same numbers a GM or controller tracks.
- Build the swarmLearned from practice runs (energy, housekeeping), or a bidding market (relocation, shifts) when a fair price can be worked out directly.
- Test on unseen dataHeld-out summer weeks, or fresh simulated nights and weeks, against today's practice, the best simple rule and the best possible plan.
- Report it plainlyThis dashboard, including where the swarm does not win yet.
What's real and what's assumed
- Real: the hourly grid demand and weather pattern behind the energy test, Ontario's Global Adjustment and hours-of-work rules, and the scheduling optimiser.
- Simulated: the five hotels, their equipment, check-outs and arrivals, oversold nights, rosters and open shifts. They are sized to be realistic and generated many times over, so a result isn't luck.
- Assumed: rates and costs, listed on each tab. They are kept in one place so they can be replaced with a property's own numbers.
To run this on a real hotel group
- Energy: the meter or BMS interval data and the equipment list per property, plus the IESO demand feed.
- Housekeeping: PMS room status, check-out times and arrivals, and the attendant roster.
- Relocation: each night's oversell position and open inventory by room type across the group.
- Staffing: the roster, wages, qualifications and open-shift requests from the scheduling system.
Each one can start as a weekly report (what the swarm would have done, and what it would have saved) before it is trusted to act.
Glossary
- Swarm
- Many small decision-makers, each acting on what it can see, coordinated by a shared goal rather than a central plan.
- Learned swarm
- A swarm whose decisions come from a model trained on practice runs (reinforcement learning). Every member uses the same model, so adding a hotel adds members, not a new model.
- Market swarm
- A swarm that coordinates through bids: each member states what a choice is worth to it, and the best overall bids win.
- Global Adjustment (GA)
- A charge on Ontario electricity bills. For Class A customers it is set by their share of the province's five peak demand hours each year.
- kW and kWh
- kW is how much power is drawn at a moment, and kWh is energy used over time. Peak charges are about kW in specific hours.
- Walking a guest
- Sending a guest with a reservation to another hotel because the booked hotel is full.
- Held-out data
- Days or scenarios kept aside and never used while building the swarm, so the test is fair.
- Best possible plan
- The mathematical optimum computed with full knowledge of the night or week in advance. It is a reference, not something a live operation can see.