Efficient and effective generation of test stimulus is key to achieving verification and validation goals. We are investigating methods to exploit machine learning for test generation in the context of microelectronic design verification. We are also researching how other AI-based techniques, such as planning, can be used for test generation.
Test Generation
Research projects:
Agent based Test Generation for Autonomous Vehicle Testing
The multi-agent systems (MAS) paradigm can be used for test generation which is an active area of research in the Trustworthy Systems Lab. Agents can be taught to find solutions to specific verification goals, such as to find pedestrian safety tests for Autonomous Vehicle testing. The video is a presentation given at the IEEE AI Test 2020 conference where we discuss using agent-based test generation to find pedestrian safety tests and show how this is more accurate than random methods.
Reinforcement Learning Gym for Test Agents
TSL researchers have developed a reinforcement learning environment specifically for generating tests for autonomous vehicle, called CAV-gym. Based on the OpenAI gym framework, this environment can be used to train agents to find interesting tests based on specific goals formulated around driving assertions, such as pedestrian safety. In the video we see an autonomous vehicle (AV) with a planned trajectory (green line) that encounters a pedestrian crossing the road causing the AV to dynamically replan it's route (blue line). In this case the pedestrian agent has generated a test which exercised the AV's collision avoidance. Take a look at our GitHub page for access to the code.