Affiliation:
1. University of Southern California Department of Aerospace and Mechanical Engineering, , 3650 McClintock Avenue, OHE 400, Los Angeles, CA 90089-1453
Abstract
Abstract
Collision avoidance in ships and robotic vehicles exemplifies a complex work process that necessitates effective scenario recognition and precise movement decision-making. Machine learning methods addressing such work processes generally involve learning from scratch, which is not only time-consuming but also demands significant computational resources. Transfer learning emerges as a potent strategy to enhance the efficiency of these engineering work processes by harnessing previously acquired knowledge from analogous tasks, thereby streamlining the learning curve for new challenges. This research delves into two critical questions central to optimizing transfer reinforcement learning for the work process of collision avoidance: (1) Which process features can be successfully transferred across varying work processes? (2) What methodologies support the efficient and effective transfer of these features? Our study employs simulation-based experiments in ship collision avoidance to address these questions, chosen for their intrinsic complexity and the varied feature recognition it demands. We investigate and compare two transfer learning techniques—feature extraction and finetuning—utilizing a lightweight convolutional neural network (CNN) model pretrained on a base case of a comparable work process. Pixel-level visual input is leveraged to cover different numbers of encountering ships and fix the input size for the model. This model adeptly demonstrates the feasibility of transferring essential features to newer work process scenarios. Further, to enhance realism and applicability, we introduce a simplified yet comprehensive ship dynamic model that considers the substantial effects of ship inertia, thereby refining the interaction between the model and its environment. The response time is embedded into the reward function design to be considered for policy training. Experimental outcomes underscore the transferability of diverse process features and evaluate the relative effectiveness of the employed transfer methods across different task settings, offering insights that could be extrapolated to other engineering work processes.
Reference47 articles.
1. Stable-Baselines3: Reliable RL Implementations;Raffin;J. Mach. Learn. Res.,2021
2. Playing Atari With Deep RL;Mnih,2013
3. Mujoco: A Physics Engine for Model-Based Control;Todorov,2012
4. Proximal Policy Optimization Algorithms;Schulman,2017
5. Ergonomics, Axiomatic Design and Complexity Theory;Suh;Theor. Issues Ergon. Sci.,2007