Dining room · 18 seats
This service
What the agent learned
Validation revenue during training. The brass and red lines are FCFS and the greedy host on the same validation nights.
The decision
Each minute the host sees the party at the head of the line and chooses one of seven actions: ask them to wait, turn them away, or seat them at a specific table. A seat is legal only when the table is free and large enough. The shift lasts 240 minutes, with arrivals slow, then heavy, then quiet.
- Tables seat 4, 2, 2, 4, and 6 guests.
- A party of two pays $40. A party of six pays $120.
- The state is 12 numbers: time left, five tables, the queue, the party, and the phase of service.
- The network is a Double DQN with two hidden layers of 64 units, γ = 0.98, and invalid seats masked out.
How to read a choice
First-come, first-served sits the party at the first table that fits. The greedy host prefers a tight fit and will turn away a small party rather than waste the six-top before late service. The DQN can do the same thing for a different reason: the Q-values say a larger check is likely before that table would free.
Keys: 1–5 seat, W wait, X turn away, Space plays an automatic host.