RoxyBot-06: Stochastic Prediction and Optimization in TAC Travel

In this paper, we describe our autonomous bidding agent, RoxyBot, who emerged victorious in the travel division of the 2006 Trading Agent Competition in a photo finish. At a high level, the design of many successful trading agents can be summarized as follows: (i) price prediction: build a model of market prices; and (ii) optimization: solve for an approximately optimal set of bids, given this model. To predict, RoxyBot builds a stochastic model of market prices by simulating simultaneous ascending auctions. To optimize, RoxyBot relies on the sample average approximation method, a stochastic optimization technique.

💡 Research Summary

The paper presents RoxyBot‑06, the autonomous bidding agent that won the travel division of the 2006 Trading Agent Competition (TAC). The authors frame the problem of automated travel procurement as a two‑stage process: (i) price prediction and (ii) bid optimization. In the price‑prediction stage, RoxyBot builds a stochastic model of market prices by simulating the simultaneous ascending auctions that govern the sale of flights, hotel rooms, and rental cars. The simulation incorporates probabilistic models of opponent behavior, auction dynamics (price increments, closing rules), and historical price trends. By running thousands of Monte‑Carlo auction scenarios, the agent derives not only expected prices but also variances and covariances across the different goods, thereby capturing the inherent uncertainty of the market.

For bid optimization, the agent employs the Sample Average Approximation (SAA) method, a well‑known stochastic optimization technique. The SAA approach replaces the original stochastic objective—maximizing expected utility given uncertain prices—with a deterministic counterpart obtained by averaging the objective over a set of sampled price scenarios generated in the prediction phase. This deterministic problem is then formulated as an integer linear program (ILP) that respects travel‑itinerary constraints (time windows, sequence of activities), budget limits, and supply constraints of each auctioned item. Solving the ILP yields a set of bids that maximizes the average expected utility across all sampled scenarios.

RoxyBot‑06 also features a real‑time feedback loop: as the actual auctions progress, observed price information is used to update the stochastic price model, which in turn refines the SAA optimization in subsequent bidding rounds. This dynamic adaptation allows the agent to correct any prediction bias and to react promptly to market shifts.

Experimental evaluation pits RoxyBot‑06 against a suite of competing TAC agents across multiple runs. Performance metrics include total revenue, client satisfaction (measured by itinerary feasibility), and cost efficiency. The results demonstrate that RoxyBot‑06 consistently outperforms rivals, achieving higher expected revenue and more stable outcomes. Notably, the inclusion of price covariances in the stochastic model and the use of multivariate SAA provide a measurable advantage over simpler mean‑price approaches. The real‑time model updating further improves prediction accuracy as the auction nears its end, leading to more precise bid placement.

The authors conclude by outlining future research directions: (1) richer opponent‑behavior models to better anticipate strategic bidding, (2) adaptive sampling strategies that balance computational effort with solution quality in large‑scale auctions, and (3) integration of reinforcement‑learning techniques to enable end‑to‑end learning of both prediction and optimization components. These extensions aim to generalize the stochastic prediction‑optimization framework of RoxyBot‑06 to broader domains such as electronic marketplaces, supply‑chain procurement, and autonomous financial trading.