A5080390843- Reinforcement Learning in Robotics
- Advanced Bandit Algorithms Research
- Supply Chain and Inventory Management
- Experimental Behavioral Economics Studies
- Model Reduction and Neural Networks
- Computability, Logic, AI Algorithms
- Protein Structure and Dynamics
- Enzyme Structure and Function
- Machine Learning in Bioinformatics
- Explainable Artificial Intelligence (XAI)
- Smart Grid Energy Management
- Misinformation and Its Impacts
Intarcia Therapeutics (United States)
2023
University of Alberta
2020-2021
Abstract Three billion years of evolution has produced a tremendous diversity protein molecules 1 , but the full potential proteins is likely to be much greater. Accessing this been challenging for both computation and experiments because space possible larger than those have functions. Here we introduce Chroma, generative model complexes that can directly sample novel structures sequences, conditioned steer process towards desired properties To enable this, diffusion respects conformational...
This paper studies model-based reinforcement learning (RL) for regret minimization. We focus on finite-horizon episodic RL where the transition model $P$ belongs to a known family of models $\mathcal{P}$, special case which is when in $\mathcal{P}$ take form linear mixtures: $P_θ = \sum_{i=1}^{d} θ_{i}P_{i}$. propose based algorithm that optimism principle: In each episode, set are `consistent' with data collected constructed. The criterion consistency total squared error incurs task...
We propose training fitted Q-iteration with log-loss (FQI-LOG) for batch reinforcement learning (RL). show that the number of samples needed to learn a near-optimal policy FQI-LOG scales accumulated cost optimal policy, which is zero in problems where acting optimally achieves goal and incurs no cost. In doing so, we provide general framework proving $\textit{small-cost}$ bounds, i.e. bounds scale achievable cost, RL. Moreover, empirically verify uses fewer than FQI trained squared loss on...
We prove that single-parameter natural exponential families with subexponential tails are self-concordant polynomial-sized parameters. For subgaussian we establish an exact characterization of the growth rate self-concordance parameter. Applying these findings to bandits allows us fill gaps in literature: show optimistic algorithms for generalized linear enjoy regret bounds both second-order (scale variance optimal arm's reward distribution) and free dependence on bound problem parameter...