To achieve intelligent behavior, agents must gather information about the world and select appropriate actions. Our hypothesis is that agents use information at different levels of abstraction to perform action selection, which is a fundamental principle of intelligence. Abstractions capture regularities in the relationship between perception and action, which can be inherent in the physics of the world or imposed via feedback control, and these regularities restrict the evolution of the system state, making action selection easier.
This thesis topic aims to identify different levels of abstractions on example 2D simulation environments, investigate their dependence on specific tasks, and explore the regularities they exploit. By understanding the regularities that exist at different levels of abstraction, an agent can choose the most effective course of action. Through your master’s thesis, you can explore how different levels of abstraction impact action selection and contribute to the evolution of an agent’s system state.
Overall, the thesis will explore how the use of abstractions and regularities can improve the performance of intelligent agents in solving different tasks. By providing a methodological underpinning for generating intelligent behavior, this thesis has the potential to contribute to the development of more advanced and effective artificial intelligence systems.
In this thesis, you will,
The scope is flexible and not limited to the listed tasks.
You can find all the necessary information here.