Publicação
Quantum-enhanced reinforcement learning
| Resumo: | The field of Artificial Intelligence has lately witnessed extraordinary results. The ability to design a system capable of beating the world champion of Go, an ancient Chinese game known as the holy grail of AI, caused a spark worldwide, making people believe that some thing revolutionary is about to happen. A different flavor of learning called Reinforcement Learning is at the core of this revolution. In parallel, we are witnessing the emergence of a new field, that of Quantum Machine Learning which has already shown promising results in supervised/unsupervised learning. In this dissertation, we reach for the interplay between Quantum Computing and Reinforcement Learning. This learning by interaction was made possible in the quantum setting using the con cept of oraculization of task environments suggested by Dunjko in 2015. In this dissertation, we extended the oracular instances previously suggested to work in more general stochastic environments. On top of this quantum agent-environment paradigm we developed a novel quantum algorithm for near-optimal decision-making based on the Reinforcement Learn ing paradigm known as Sparse Sampling, obtaining a quantum speedup compared to the classical counterpart. The achievement was a quantum algorithm that exhibits a complexity independent on the number of states of the environment. This independence guarantees its suitability for dealing with large state spaces where planning may be inapplicable. The most important open questions remain whether it is possible to improve the orac ular instances of task environments to deal with even more general environments, especially the ability to represent negative rewards as a natural mechanism for negative feedback instead of some normalization of the reward and the extension of the algorithm to perform an informed tree-based search instead of the uninformed search proposed. Improvements on this result would allow the comparison between the algorithm and more recent classical Reinforcement Learning algorithms. |
|---|---|
| Autores principais: | Sequeira, André Manuel Resende |
| Assunto: | Quantum computation Sparse sampling Quantum reinforcement learning Computação quântica Amostragem esparsa Aprendizagem por reforço quântica Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática |
| Ano: | 2021 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | The field of Artificial Intelligence has lately witnessed extraordinary results. The ability to design a system capable of beating the world champion of Go, an ancient Chinese game known as the holy grail of AI, caused a spark worldwide, making people believe that some thing revolutionary is about to happen. A different flavor of learning called Reinforcement Learning is at the core of this revolution. In parallel, we are witnessing the emergence of a new field, that of Quantum Machine Learning which has already shown promising results in supervised/unsupervised learning. In this dissertation, we reach for the interplay between Quantum Computing and Reinforcement Learning. This learning by interaction was made possible in the quantum setting using the con cept of oraculization of task environments suggested by Dunjko in 2015. In this dissertation, we extended the oracular instances previously suggested to work in more general stochastic environments. On top of this quantum agent-environment paradigm we developed a novel quantum algorithm for near-optimal decision-making based on the Reinforcement Learn ing paradigm known as Sparse Sampling, obtaining a quantum speedup compared to the classical counterpart. The achievement was a quantum algorithm that exhibits a complexity independent on the number of states of the environment. This independence guarantees its suitability for dealing with large state spaces where planning may be inapplicable. The most important open questions remain whether it is possible to improve the orac ular instances of task environments to deal with even more general environments, especially the ability to represent negative rewards as a natural mechanism for negative feedback instead of some normalization of the reward and the extension of the algorithm to perform an informed tree-based search instead of the uninformed search proposed. Improvements on this result would allow the comparison between the algorithm and more recent classical Reinforcement Learning algorithms. |
|---|
Atividades financiadas
Não foram encontrados projetos financiados.