| Abstract: |  Presented Diploma Thesis is from area of development of artificial creatures. The work builds on a research of the latest approaches to artificial life, realized by the Department of Cybernetics, CTU in Prague, under the leadership of Pavel Nahodil in the last twenty years.
The main feature of this architecture should be it's total autonomy, ability to gain all information from the surrounding environment and effective information filtering and classification. Agent is supposed to operate based only on the sensory input and by it's actuator system, because of the properties meant above the resulting architecture should be almost fully independent on the concrete area and form of use. Consequently it is unimportant whether the agent is embodied in some robotic system, intelligent house, or just operates in some virtual environment. Thanks to the fact that all the designer has to specify is just the sensory layer, actuator layer, and agent's needs, the architecture should be convenient especially in unknown environments, where some more complex task has to be fulfilled.
This agent architecture is based on the layered model, combining various approaches on different layers. The life of agent is similar to a newly born animal, which explores new and unknown environment, learns from experiences and links the newly learned abilities to it's needs in order to survive and increase effectiveness of it's behavior. New knowledge is learned simultaneously on various levels of abstraction using different learning approaches.
Besides the proposing the agent architecture, I hope that this thesis should be contribution also in the approach of combining some various types of decision making, the resulting system exploits their benefits and inhibits their weaknesses. One of the main features is an alternative implementation of system similar to reactive and hierarchical planning. The system combines hierarchical reinforcement learning with planning engine. The main benefit should be the domain independency of this planner, because the entire hierarchy of actions is created completely autonomously, the other benefit is tight connection of these two systems, so the boundary how big part of given task should be created deliberatively and which rather reactively (based on the concrete actual situation), can be chosen online based on the specific agent's preferences. That's why I hope that the proposed part of architecture could be used also as an advanced domain independent hierarchical planning engine with autonomous learning from the experiences.
|
|---|