Technical Reports
TR001 PCT Applied to Lunar Lander with Comparative RL Baseline TR001_v1.0.12.pdf
This research investigates the comparative performance of Perceptual Control Theory (PCT) and Reinforcement Learning (RL) within the Lunar Lander simulation environment, aiming to address the research gap regarding the efficacy of PCT as a control strategy for autonomous agents.
Key findings reveal that the PCT controller achieved a higher success rate of 79% compared to the RL’s 75%, despite utilising significantly fewer parameters by a factor of 10,000 (29 versus 335,622). This result underscores PCT’s computational efficiency and robustness, making it a viable alternative in scenarios where resource constraints and system transparency are critical.
TR002 PCT Controller solution: MountainCarContinuous-v0: Description and Analysis MountainCarContinuous-v0_best_controller_report.pdf
This report examines a verified hierarchical Perceptual Control Theory controller for Gymnasium’s MountainCarContinuous-v0 task. It explains the controller’s position-control and velocity-control hierarchy, its feedback signals and action reversal, and the resulting momentum-building behaviour.
Across 100 unique reset seeds, the controller completed 100 of 100 episodes, taking 76–93 steps and achieving total reward from 93.520 to 95.756. The report also links to a replay video showing the final controller and its progressive development through the evolutionary process.