Hierarchical intrinsically motivated agent planning behavior with dreaming in grid environments

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Reinforcement Learning

Background:

• Biologically plausible learning models are essential for developing advanced autonomous intelligent agents.
• Current agents often struggle in unfamiliar environments and require extrinsic reinforcement signals.
• Hierarchical reinforcement learning offers a framework for complex task performance.

Purpose of the Study:

• To propose a novel hierarchical agent architecture inspired by biological learning mechanisms.
• To enable autonomous agents to learn and act effectively in unknown environments.
• To investigate the role of intrinsic motivation and simulated experience (dreaming) in agent learning.

Main Methods:

• A hierarchical agent model incorporating temporal memory for state-action representation.
• Utilizing a basal ganglia model for learning action policies at multiple abstraction levels.
• Implementing an intrinsic motivation system driven by the agent's learned environment model.
• Employing a 'dreaming' mechanism for acting in imagination.

Main Results:

• The proposed agent architecture demonstrated effective goal achievement in grid environments.
• Temporal memory facilitated sparse distributed representations of state-actions.
• The basal ganglia model supported hierarchical policy learning.
• Intrinsic motivation successfully guided the agent in the absence of extrinsic signals.

Conclusions:

• The developed hierarchical model provides a biologically plausible approach to autonomous intelligent agents.
• Intrinsic motivation and simulated experience are effective mechanisms for exploration and learning.
• The architecture shows promise for creating agents capable of complex, adaptive behaviors in diverse environments.