Abstract
Respiratory droplets are considered the main transmission mode of the SARS-CoV-2 virus, so social distancing has been widely adopted as a countermeasure. Individuals from different households are expected to keep a physical distance, while most public venues, such as restaurants, theaters and cinemas, are shut down completely resulting in undesirable economic and social effects. Lately, the unmatched development and application of vaccines have paved a path for reopening, but uncertainties surrounding virus mutations and vaccination rates may require to retain basic physical distancing measures for this transition period.
One of the challenges to reopening public venues in a safe and economically feasible manner is seat allocation, i.e. maximizing the number of occupiable seats while respecting distancing requirements between groups. Static layouts can efficiently be predetermined with classic optimization methods, while dynamic layouts offer more flexibility with respect to the arriving group sizes, but are also more difficult to determine.
In our work, we address the dynamic seat allocation problem with reinforcement learning (RL) -- a paradigm in machine learning of training a model via its interaction with an environment. In this lightning talk, we will share our intermediate results and experiences from this problem: first, how to model the problem and the learning environment (formulating state and action spaces, reward function, group size distribution, distancing constraints, group closeness) and, second, suitability, potentials and limitations of both classic and state-of-the-art RL methods (Q-learning, proximal policy optimization, tree search methods).
One of the challenges to reopening public venues in a safe and economically feasible manner is seat allocation, i.e. maximizing the number of occupiable seats while respecting distancing requirements between groups. Static layouts can efficiently be predetermined with classic optimization methods, while dynamic layouts offer more flexibility with respect to the arriving group sizes, but are also more difficult to determine.
In our work, we address the dynamic seat allocation problem with reinforcement learning (RL) -- a paradigm in machine learning of training a model via its interaction with an environment. In this lightning talk, we will share our intermediate results and experiences from this problem: first, how to model the problem and the learning environment (formulating state and action spaces, reward function, group size distribution, distancing constraints, group closeness) and, second, suitability, potentials and limitations of both classic and state-of-the-art RL methods (Q-learning, proximal policy optimization, tree search methods).