Bounded Activation Function for Autonomous Driving Reliability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for autonomous driving using reinforcement learning struggle to incorporate reliability constraints, leading to complex regulation scenarios and high computational effort, especially in limited automotive computing resources, and require pre-identification of undesired actions.
Innovation Solution
A method utilizing a deep neural network with a bounded activation function and bound extraction function to limit vehicle movements within predefined state-dependent bounds, eliminating the need for artificial reliability constraints and manual prioritization processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If reinforcement learning is used for autonomous driving, then the virtual driver can automatically gain optimal behavior through environmental interactions, but reliability constraints are difficult to incorporate and may lead to unforeseen configurations during training
Solution Approach 1:
The patent introduces a bounded activation function as an intermediary component between the neural network and the action space. This mediator automatically enforces reliability constraints by bounding the neural network's output activations, thereby preventing unforeseen configurations without requiring manual constraint specification or complex regulatory layers.
Solution Approach 2:
The patent applies preliminary action by pre-defining bounds for the activation function before the reinforcement learning training begins. These bounds are hard-coded into the neural network architecture, ensuring that reliability constraints are satisfied from the outset rather than being added as post-processing corrections during training.
2Reliability
If a specified controller is included to monitor and convert actions into reliable actions, then reliability is improved, but the complexity of regulation scenarios increases due to considering all possible action spaces
Solution Approach 1:
The patent extracts the reliability constraint enforcement mechanism from the complex regulatory controller and embeds it directly into the neural network's activation function. This eliminates the need for separate monitoring and conversion processes, reducing regulation complexity while maintaining reliability through the bounded activation function that inherently limits outputs to safe ranges.
3Reliability
If an additional optimization procedure is applied at every timestep to optimize actions with reliability conditions, then reliability is improved, but computational effort becomes too high for automotive applications
Solution Approach 1:
The patent merges the reliability constraint enforcement with the neural network's forward propagation process by implementing bounds directly in the activation function. This combination eliminates the need for separate optimization procedures at each timestep, as reliability is ensured during the standard neural network computation, thereby reducing computational effort to levels suitable for automotive applications.
4Reliability
If the available action set is pruned to exclude undesired actions, then reliability is improved, but computational resources are consumed in pre-processing or post-processing steps
Solution Approach 1:
The patent applies preliminary action by pre-defining bounds for the activation function that inherently exclude undesired actions. By encoding these bounds into the neural network architecture before training, the system prevents invalid actions from being generated in the first place, eliminating the need for computationally expensive post-processing filtering or pre-processing action space pruning.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure generally relates to a method for autonomous driving of a vehicle, a data processing circuit, a computer program, and a computer-readable medium. The vehicle comprises at least a sensing device and at least one movement controller configured for influencing a movement of the vehicle if at least one control value is provided to the movement controller. The method comprises at least determining at least one state of the vehicle based on data acquired, with a sensing device, with respect to an environment of the vehicle. The method also comprises determining, using a deep neural network, at least one normalized action for the at least one movement controller of the vehicle based on a bounded activation function. Also, the method comprises mapping the at least one normalized action to at least one control value using a bound extraction function.