Defining a Metric-Driven Approach for Learning Hazardous Situations

Artificial intelligence has brought many innovations to our lives. At the same time, it is worth designing robust safety machine learning (ML) algorithms to obtain more benefits from technology. Reinforcement learning (RL) being an important ML method is largely applied in safety-centric scenarios. In such a situation, learning safety constraints are necessary to avoid undesired outcomes. Within the traditional RL paradigm, agents typically focus on identifying states associated with high rewards to maximize its long-term returns. This prioritization can lead to a neglect of potentially hazardous situations. Particularly, the exploration phase can pose significant risks, as it necessitates actions that may have unpredictable consequences. For instance, in autonomous driving applications, an RL agent might discover routes that yield high efficiency but fail to account for sudden hazardous conditions such as sharp turns or pedestrian crossings, potentially leading to catastrophic failures. Ensuring the safety of agents operating in unpredictable environments with potentially catastrophic failure states remains a critical challenge. This paper introduces a novel metric-driven approach aimed at containing risk in RL applications. Central to this approach are two developed indicators: the Hazard Indicator and the Risk Indicator. These metrics are designed to evaluate the safety of an environment by quantifying the likelihood of transitioning from safe states to failure states and assessing the associated risks. The fact that these indicators are characterized by a straightforward implementation, a highly generalizable probabilistic mathematical foundation, and a domain-independent nature makes them particularly interesting. To demonstrate their efficacy, we conducted experiments across various use cases, showcasing the feasibility of our proposed metrics. By enabling RL agents to effectively manage hazardous states, this approach paves the way for a more reliable and readily implementable RL in practical applications.