| Title | : | A Thompson Sampling Approach to Linearly Constrained Structured Link Adaptation |
| Speaker | : | Utsav Dey (IITM) |
| Details | : | Wed, 5 Apr, 2023 11:00 AM @ SSB 223 |
| Abstract: | : | We consider the problem of learning to choose encoders in a sequential fashion where the encoded data is transmitted via a stochastically chosen channel by the environment .After every round of transmission as feedback, the algorithm obtains a binary response indicating whether the transmission was successful or not.We model this using the multi-armed bandit framework where the arms correspond to the encoders. Instead of the standard notion of regret, we measure algorithmic performance using a more natural Rate to Violation (R2V) ratio. We propose a Thompson Sampling-based algorithm called LinConErrTS(and a neural network based variant ) to solve this problem. We provide theoretical insights into how the rank of a particular error matrix affects the performance of our algorithm and also show a sample complexity result for bounded per round sub-optimality. We experiment our algorithm under several settings and show that it outperforms several baseline algorithms. |
