2. Revenue Management System (RMS)
Introduction
− RMS: automated system
− Goal: maximize flight revenue
− All seats
− Whole reservation period (1 year)
− What the systems does:
− Set the right fares at the right time
3. Motivation – limitations of RMS
RMS assumptions
− Rely of historical observations – may represent the future or not
− Collection of historical observations
− The future is similar to the past (e.g. schedules, fares, volume, WTP, arrival rate, etc.)
− Rely on a model representation – may be right or wrong
− Requires a demand model for volume and WTP
− Requires an optimization model (e.g. EMSR or Dynamic Programming)
− No exploration – no validation of assumptions
− RMS always takes the “optimal” decision
− Assumes monopoly – in reality there is almost always competition
− Average policies over all competitors states and actions
17. Performance of RMS vs. RMS; DQL vs. RMS; DQL vs. DQL
Revenue wrt baseline
Booking Class Mix
Price evolution
5,36%
-8,32%
-15%
-10%
-5%
0%
5%
10%
DQL vs. RMS DQL vs. DQL
50
100
150
200
0 5 10 15 20
RMS vs RMS DQL vs. RMS DQL vs. DQL
0
5
10
15
1 2 3 4 5 6 7 8 9 10
RMS vs RMS
0
5
10
15
1 2 3 4 5 6 7 8 9 10
DQL vs. RMS
0
5
10
15
1 2 3 4 5 6 7 8 9 10
DQL vs. DQL
RMS vs. RMS
DCP
Class
Pax
18. 18
Conclusion
− Reinforcement Learning (RL) opens the door to a radical new approach
• Model free - no forecasting and no optimization
• Leans by direct price testing
− Shown that QL and DQL converges to RMS optimal solution in monopoly
− Why does DQL vs. DQL underperform?
• Hypothesis: Both ALs agree that WTP is increasing in traditional RMS?
− Next steps
• Apply to networks
• Add more information to the state (e.g., competitors and market)
*) Nash equilibrium
RMS DQL
RMS 100; 100 84.2; 105.4
DQL 105.4; 84.2 91,7*; 91.7*