Exploration Strategies for Homeostatic Agents: Continuous and dynamic exploration for homeostatic regulation using deep reinforcement learning
Typ
Examensarbete för masterexamen
Program
Publicerad
2019
Författare
Andersson, Patrick
Strandman, Anton
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
This paper introduces and evaluates four novel exploration strategies for homeostatic
agents. Homeostatic agents have the objective of keeping some internal variables
as close to a predetermined optimum as possible. Reinforcement learning is used
for decision making, and the agents are given access to the optimal and acceptable
values of the internal variables, giving greater flexibility for exploration and better
survival chances. The new exploration strategies that utilise the internal variables
are evaluated in a range of environments, showing them to outperform common
reinforcement learning exploration techniques where these variables are not taken
into consideration.
Beskrivning
Ämne/nyckelord
artificial general intelligence , multi-objective reinforcement learning , exploration , homeostatic regulation , animat , homeostatic exploration