Toward a Brain-Inspired System: Deep Recurrent Reinforcement Learning for a Simulated Self-Driving Agent.

Area of Science:

• Artificial Intelligence
• Neuroscience
• Robotics

Background:

• Bio-inspired learning methods offer alternatives to classical programming.
• Reinforcement learning (RL) shows strong links to neural substrates, driving interest in decision-making tasks.
• Simulating human-like autonomous control in robots, particularly for self-driving, is a key research area.

Purpose of the Study:

• To propose an end-to-end deep-Q network architecture with recurrence for simulated self-driving.
• To train a driving agent using a brain-inspired trial-and-error technique mirroring real-world learning.
• To address limitations in current deep learning approaches for autonomous driving.

Main Methods:

• Developed a novel deep-Q network architecture incorporating recurrence.
• Utilized raw screen outputs as the sole input for the driving agent.
• Implemented a weighted layer to enhance long episode differences and a modified replay mechanism to combat sparsity and accelerate learning.
• Trained and tested the network within an OpenAI Gym environment.

Main Results:

• The trained driving agent demonstrated advanced autonomous behaviors in simulated driving scenarios.
• The brain-inspired trial-and-error approach proved effective in learning complex driving tasks.
• The proposed network innovations successfully addressed challenges like data sparsity and learning speed.

Conclusions:

• The developed brain-inspired learning system offers a promising approach for practical self-driving control solutions.
• This study highlights the potential of integrating neuroscience principles with deep reinforcement learning for robotics.
• Further research can build upon this architecture to enhance the capabilities of autonomous agents.