Publicação
Quantum bayesian reinforcement learning
| Resumo: | Reinforcement learning has had many recent achievements and is becoming increasingly more relevant in the scientific community. As such, this work uses quantum computing to find potential advantages over classical reinforcement learning algorithms, using Bayesian networks to model the considered decision making environments. For this purpose, this work makes use of quantum rejection sampling, a quantum approximate inference algorithm for Bayesian networks proposed by Low et al. [2014] with a quadratic speedup over its classical counterpart for sparse networks. It is shown that this algorithm can only provide quantum speedups for partially observable environments, and a quantum-classical hybrid lookahead al gorithm is presented to solve these kinds of problems. Moreover, this work also includes both sample and computational complexity analysis of both this quantum lookahead algorithm and its classical alternative. While the sample complexity is shown to be identical for both algorithms, the quantum approach provides up to a quadratic speedup in computational complexity. Finally, the potential advantages of this new algo rithm are experimentally tested in different small experiments. The results show that this speedup can be leveraged either to improve the rational decision-making skills of agents or to reduce their decision-making time due to the reduction in computational complexity. |
|---|---|
| Autores principais: | Cunha, Gilberto Rui Nogueira |
| Assunto: | Reinforcement learning Bayesian networks Quantum computing Quantum decision-making Aprendizagem por reforço Redes Bayesianas Computação quântica Tomada de decisão quântica |
| Ano: | 2023 |
| 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: | Reinforcement learning has had many recent achievements and is becoming increasingly more relevant in the scientific community. As such, this work uses quantum computing to find potential advantages over classical reinforcement learning algorithms, using Bayesian networks to model the considered decision making environments. For this purpose, this work makes use of quantum rejection sampling, a quantum approximate inference algorithm for Bayesian networks proposed by Low et al. [2014] with a quadratic speedup over its classical counterpart for sparse networks. It is shown that this algorithm can only provide quantum speedups for partially observable environments, and a quantum-classical hybrid lookahead al gorithm is presented to solve these kinds of problems. Moreover, this work also includes both sample and computational complexity analysis of both this quantum lookahead algorithm and its classical alternative. While the sample complexity is shown to be identical for both algorithms, the quantum approach provides up to a quadratic speedup in computational complexity. Finally, the potential advantages of this new algo rithm are experimentally tested in different small experiments. The results show that this speedup can be leveraged either to improve the rational decision-making skills of agents or to reduce their decision-making time due to the reduction in computational complexity. |
|---|