#### DMCA

## Adaptive mechanism design: a metalearning approach (2006)

### Cached

### Download Links

- [www.cs.utexas.edu]
- [www.cs.utexas.edu]
- [www.cs.utexas.edu]
- [www.cs.utexas.edu]
- [www.cs.utexas.edu]
- [www.cs.utexas.edu]
- [www.cs.utexas.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | In Proceedings of the Eighth International Conference on Electronic Commerce (ICEC ’06 |

Citations: | 11 - 1 self |

### Citations

6309 | Prospect Theory: An Analysis of Decision Under
- Kahneman, Tversky
- 1979
(Show Context)
Citation Context ...ctions, this inefficiency is a serious drawback. Perhaps the biggest challenge results from the fact that, in practice, bidders are not able to attain full rationality in complex, real-world settings =-=[9]-=-. Rather, they employ heuristic strategies that are in general opaque to the seller, certainly a priori, and often even after the auction. One method of addressing these challenges that has received r... |

5612 | Reinforcement Learning: An Introduction,
- Sutton, Barto
- 1998
(Show Context)
Citation Context ... reserve price at each step, where the ith choice is a price of (i−1)/(k−1). The resulting problem can be viewed as an instance of the k-armed bandit problem, a classic reinforcement learning problem =-=[18]-=-. In karmed bandit problems, the expected value of each choice is assumed to be independent, and the goal of maximizing the reward obtained presents a tradeoff between exploring the choices, in order ... |

1605 |
Optimal Auction Design
- Myerson
- 1981
(Show Context)
Citation Context ...e of any bid higher than the reserve price, no transaction occurs. It has been shown that when bidders are rational, the optimal reserve price should be higher than the seller’s valuation of the item =-=[10]-=-; however, a reserve price of 0 is often seen in practice. Dodonova and Khoroshilov explain this phenomenon by bidders’ loss aversion [8]. Loss aversion violates the rationality assumption because the... |

599 | Locally weighted learning
- Atkeson, Moore, et al.
- 1997
(Show Context)
Citation Context ...of ongoing work.sfollows. For each episode the seller wishes to simulate, it first randomly generates an “arbitrary” distribution for valuations by taking a Gaussian with a mean chosen uniformly from =-=[0, 1]-=- and a variance of 10 x with x chosen uniformly from [-2, 1], and then normalizes the distribution so that the portion over the range [0, 1] represents a PDF. The seller then generates a distribution ... |

158 | Iterative Combinatorial Auctions: Achieving Economic and Computational Efficiency
- Parkes
- 2001
(Show Context)
Citation Context ...nism design has traditionally been largely an analytic process. Assumptions such as full rationality are made about bidders, and the resulting properties of the mechanism are analyzed in this context =-=[12]-=-. Even in large-scale realworld auction settings such as the FCC Spectrum auctions, game theorists have convened prior to the auction to determine the best mechanism to satisfy a set of objectives. Hi... |

157 | The FCC spectrum auctions: An early assessment
- Cramton
- 1997
(Show Context)
Citation Context ...o determine the best mechanism to satisfy a set of objectives. Historically, this process has been incremental, requiring several live iterations to iron out wrinkles, and the results have been mixed =-=[6, 20]-=-. An important component of this incremental design process involves reevaluating the assumptions made about bidders in light of auction outcomes. In particular, these assumptions pertain to bidders’ ... |

117 | A perspective view and survey of meta-learning.
- Vilalta, Drissi
- 2002
(Show Context)
Citation Context ...y, we will use the term encountered population in the following text whenever referring to the actual encountered population, and not a simulated population.) In this paper, we explore a metalearning =-=[19]-=- approach to choosing the adaptive algorithm. Specifically, we choose an adaptive algorithm that is itself parameterized, and then search for the parameters that result in the best performance under e... |

115 | A parameterization of the auction design space
- Wurman, Wellman, et al.
(Show Context)
Citation Context ...rameters may be considered, such as reserve prices, auctioneer fees, minimum bid increments, and whether the close is hard or soft. (For an extensive parameterization of the auction design space, see =-=[21]-=-.) The adaptive algorithm is essentially an online machine learning algorithm aiming to characterize the function from mechanism parameters to expected revenue (or any other objective function). Becau... |

90 | An overview of the simultaneous perturbation method for efficient optimization,”
- Spall
- 1998
(Show Context)
Citation Context ...e large number of random factors involved in the process, and so we are faced with a stochastic optimization task. To solve this task, we use Simultaneous Perturbation Stochastic Approximation (SPSA) =-=[17]-=-, a method of stochastic optimization based on gradient approximation. At each step, two estimates of the expected episode revenue are taken for slight perturbations of the current parameters (the sam... |

70 | Online learning in online auctions.
- Blum, Kumar, et al.
- 2003
(Show Context)
Citation Context ...of addressing these challenges that has received recent attention is the use of machine learning algorithms to revise auction parameters in response to observed bidder behavior. For instance, [2] and =-=[3]-=- consider the problem of maximizing seller revenue in online auctions throughsthe use of online learning algorithms for combining “expert” advice. Such approaches differ significantly from the traditi... |

64 |
Making more from less: Strategic demand reduction in the FCC spectrum auctions
- Weber
- 1997
(Show Context)
Citation Context ...o determine the best mechanism to satisfy a set of objectives. Historically, this process has been incremental, requiring several live iterations to iron out wrinkles, and the results have been mixed =-=[6, 20]-=-. An important component of this incremental design process involves reevaluating the assumptions made about bidders in light of auction outcomes. In particular, these assumptions pertain to bidders’ ... |

57 | Evolution of market mechanism through a continuous space of auction-types.
- Cliff
(Show Context)
Citation Context ...chanisms in response to bidder behavior from an empirical standpoint using a variety of learning approaches in simulation. In this section, we briefly survey that work and relate it to our own. Cliff =-=[5]-=- explores a continuous space of auction mechanisms defined by a parameterized continuous double auction, where the parameter represents the probability that a seller will make an offer during any time... |

49 | Near-optimal online auctions
- Blum, Hartline
- 2005
(Show Context)
Citation Context ... method of addressing these challenges that has received recent attention is the use of machine learning algorithms to revise auction parameters in response to observed bidder behavior. For instance, =-=[2]-=- and [3] consider the problem of maximizing seller revenue in online auctions throughsthe use of online learning algorithms for combining “expert” advice. Such approaches differ significantly from the... |

44 | Active learning for class probability estimation and ranking. In:
- SaarTsechansky, Provost
- 2001
(Show Context)
Citation Context ... the seller can select its own training examples (by choosing which set of auction parameters to try next), and because the target output is, in general, continuous, the problem is an active learning =-=[15]-=- regression problem. A key characteristic is that the learning is all done online, so that excessive exploration can be costly. The bidders in Figure 1 may use a variety of different bidding strategie... |

41 | Co-evolutionary auction mechanism design
- Phelps, Burnley, et al.
- 2002
(Show Context)
Citation Context ...ake sense, the same bidders must interact repeatedly with the mechanism, leading to a potential co-evolutionary scenario in which the bidders and mechanism continue to adapt in response to each other =-=[13]-=-. However, our approach does not depend on repeated interactions with the same bidders. The only required assumption about the bidders is that their behavior is somewhat consistent (e.g. bidders assoc... |

38 | Applying evolutionary game theory to auction mechanism design.
- Byde
- 2003
(Show Context)
Citation Context ...nt values of the auction parameter. Phelps et al. [13] also address continuous double auctions, using genetic programming to co-evolve buyer and seller strategies and auction rules from scratch. Byde =-=[4]-=- takes a similar approach in studying the space of auction mechanisms between the first and second-price sealed-bid auction. The winner’s payment is determined as a weighted average of the two highest... |

19 |
Exploring Auction Databases through Interactive Visualization”, Decision Support Systems
- Shmueli, Jank, et al.
- 2006
(Show Context)
Citation Context ...xtensive historical data on past auctions of identical items is available (as is the case with eBay), it may be possible for the seller to estimate the optimal parameters by analyzing this data (e.g. =-=[16]-=-). This approach is not always possible, however. If the seller is introducing a new item to the market, no such data will be available. Alternatively, if there is a sudden change in demand for an ite... |

10 |
Loss Aversion and Learning to Bid;
- Dennis, Guth, et al.
- 2005
(Show Context)
Citation Context ...iscrete bid levels in an English auction, and the bidder parameters to be estimated based on auction outcomes are the number of bidders participating and their valuation distribution. Dittrich et al. =-=[7]-=- present a different take on adaptation involving loss averse bidders, analyzing the effect that loss aversion has on the learning dynamics exhibited by bidders adapting in response to experience. 7. ... |

9 | 2006) Learning environmental parameters for the design of optimal English auctions with discrete bid levels
- Rogers, David, et al.
(Show Context)
Citation Context ...safe to use and capable of quickly finding the parameters best suited to the participating bidders, all while making as few assumptions as necessary about the behavior of these bidders. Rogers et al. =-=[14]-=- provide an example of using Bayesian inference to determine optimal auction parameters. As discussed previously, such an approach is suitable when it is known that the behavior of bidders can be full... |

6 |
Y.: Optimal auction design when bidders are loss averse. Working Paper
- Dodonova, Khoroshilov
- 2004
(Show Context)
Citation Context ...rice should be higher than the seller’s valuation of the item [10]; however, a reserve price of 0 is often seen in practice. Dodonova and Khoroshilov explain this phenomenon by bidders’ loss aversion =-=[8]-=-. Loss aversion violates the rationality assumption because the utility from a gain is lower than the disutility from a loss of the same magnitude. Specifically, if the marginal utility from winning a... |

4 |
Developing adaptive auction mechanisms. SIGecom Exchanges
- Pardoe, Stone
- 2005
(Show Context)
Citation Context ... sense to choose an adaptive algorithm that would work well 1 This scenario suggests the possibility of selling multiple items simultaneously, a topic explored in previous work on adaptive mechanisms =-=[11]-=-. In this paper, however, we restrict our attention to sequential, single item auctions, which may be most appropriate or even required in some settings. if this information is correct. The seller cou... |