Abstract
Problem: An Online Collaborative AI System (OL-CAIS) learns online from human collaboration to achieve a common goal. It may be subjected to environmental events that disrupt the system’s decision-making policies, resulting in performance degradation. Decision-makers need to develop appropriate policies to restore performance from degradation (i.e., ensure resilience) while managing the energy’s adverse effects that may result from these policies (i.e., ensure greenness). These policies should aim to select actions within OL-CAISs that lead to green recovery. Methodology: In this paper, we introduce the GResilience framework, which automatically recommends OL-CAIS actions to balance greenness and resilience. It quantifies the actions’ contribution to the two system’s qualities and then solves optimization and game theory to propose one-agent and two-agent policies, respectively. The internal policies and the GResilience ones are implemented and evaluated in three lab experiments on a real-world OL-CAIS experiencing performance degradation. In each experiment, we model system resilience and use the models with metrics to evaluate the overall effectiveness of the policies for greenness and resilience. Results: For resilience, two-agent policies achieved faster recovery and steadier performance than internal policies. Although one-agent improved performance steadiness, the system did not recover. As for greenness, both GResilience policies halved the dependency on human interactions. However, this improvement comes with a marginal increase in CO2 emissions. Conclusion: Our results demonstrate the effectiveness of the GResilience framework, emphasizing that two-agent policies provide a more balanced trade-off between greenness and resilience compared to the other policies.