Skip to main content Skip to main navigation

Project | QuBer-KI

Duration:

Quantum Deep Reinforcement Learning for simple robotic behaviours

Quantum machine learning, in particular quantum deep reinforcement learning (QDRL), has gained tremendous attention in the scientific community in recent years, as promising results for simple tasks in simple simulation environments have been found quickly. The algorithms published so far have the advantage that the architecture of the quantum circuit is such that it can be executed for small but relevant examples on currently available quantum hardware. A number of open scientific questions arise from the relative novelty of the hybrid QDRL methods. First, the reproducibility and transferability of the published approaches is unclear. Furthermore, especially the scalability and thus the possible application in more complex learning scenarios needs to be investigated.

For this purpose, however, possibilities for encoding high-dimensional input data into quantum circuits in particular need to be evaluated. Furthermore, it has to be clarified which of the numerous currently successful classical algorithms for reinforcement learning can be usefully extended by components of quantum machine learning. In addition, the investigation and development of new original hybrid QDRL algorithms are fields with significant potentials, since here methods can be developed specifically, which in their structure already provide for the possibilities of quantum algorithms.

The QuBER-KI project will make a significant contribution to answering these questions. To this end, the project will develop structured answers to these questions by first structuring, identifying and modularizing the current state of research on quantum-enhanced algorithms for reinforcement learning. Based on this, the complexity of the environments will be increased in order to determine to what extent the existing algorithms can be transferred to other environments and applications. On the other hand, the neural networks of the classical algorithms currently successfully used in practice for reinforcement learning will be replaced with a quantum circuit. This will create a module in which different algorithms with quantum circuitry can be tested in different environments. This will provide a comprehensive picture of which algorithms achieve the best results for which learning problems.

Sponsors

BMWK - Federal Ministry for Economic Affairs and Climate Action

BMWK - Federal Ministry for Economic Affairs and Climate Action