Online learning and decision-making from implicit feedback
| dc.contributor.advisor | Shakkottai, Sanjay | |
| dc.contributor.advisor | Vishwanath, Sriram | |
| dc.contributor.committeeMember | Baccelli, Francois | |
| dc.contributor.committeeMember | Zitkovic, Gordon | |
| dc.contributor.committeeMember | Srikant, Rayadurgam | |
| dc.creator | Krishnasamy, Subhashini | |
| dc.date.accessioned | 2017-06-20T20:33:42Z | |
| dc.date.available | 2017-06-20T20:33:42Z | |
| dc.date.issued | 2017-05 | |
| dc.date.submitted | May 2017 | |
| dc.date.updated | 2017-06-20T20:33:43Z | |
| dc.description.abstract | This thesis focuses on designing learning and control algorithms for emerging resource allocation platforms like recommender systems, 5G wireless networks, and online marketplaces. These systems have an environment which is only partially known. Thus, the controllers need to make resource allocation decisions based on implicit feedback obtained from the environment based on past actions. The goal is to sequentially select actions using incremental feedback so as to optimize performance while simultaneously learning about the environment. We study three problems which exemplify this setting. The first is an inference problem which requires identification of sponsored content in recommender systems. Specifically, we ask if it is possible to detect the existence of sponsored content disguised as genuine recommendations using implicit feedback from a subset of users of the recommender system. The second problem is the design of scheduling algorithms for switch networks when the user-server link statistics are unknown (for e.g., in wireless networks, online marketplaces). The scheduling algorithm has to tradeoff between scheduling the optimal links and obtaining sufficient feedback about all the links for accurate estimates. We observe the close connection of this problem to the stochastic multi-armed bandit problem and analyze bandit-style explore-exploit algorithms for learning the statistical parameters while simultaneously assigning servers to users. The third is the joint problem of base station activation and rate allocation in an energy efficient wireless network when the channel statistics are unknown. The controller observes instantaneous channel rates of activated BSs, and thereby sequentially obtains implicit feedback about the channel. Here again, there is a tradeoff between learning the channel versus optimizing the operation cost based on estimated parameters. For each of these systems, we propose algorithms with provable asymptotic guarantees. These learning algorithms highlight the use of implicit feedback in online decision making and control. | |
| dc.description.department | Electrical and Computer Engineering | |
| dc.format.mimetype | application/pdf | |
| dc.identifier | doi:10.15781/T2P26Q88C | |
| dc.identifier.uri | http://hdl.handle.net/2152/47285 | |
| dc.language.iso | en | |
| dc.subject | Online learning | |
| dc.subject | Resource allocation | |
| dc.subject | Learning algorithm design | |
| dc.subject | Implicit feedback | |
| dc.subject | Incremental feedback | |
| dc.subject | Sponsored content | |
| dc.subject | Sponsored content detection | |
| dc.subject | Scheduling algorithm design | |
| dc.subject | Stochastic multi-armed bandit | |
| dc.subject | Base station activation | |
| dc.subject | Learning algorithms | |
| dc.subject | Online decision-making | |
| dc.subject | Online control | |
| dc.title | Online learning and decision-making from implicit feedback | |
| dc.type | Thesis | |
| dc.type.material | text | |
| thesis.degree.department | Electrical and Computer Engineering | |
| thesis.degree.discipline | Electrical and Computer Engineering | |
| thesis.degree.grantor | The University of Texas at Austin | |
| thesis.degree.level | Doctoral | |
| thesis.degree.name | Doctor of Philosophy |