Periodic Reporting for period 1 - CREXDATA (Critical Action Planning over Extreme-Scale Data)
Reporting period: 2023-01-01 to 2024-06-30
The vision of CREXDATA is to develop a generic platform for real-time critical situation management including flexible action planning and agile decision making over streaming data of extreme scale and complexity. CREXDATA develops the algorithmic apparatus, software architectures and tools for federated predictive analytics and forecasting under uncertainty. The envisioned framework boosts proactive decision making providing highly accurate and transparent short- and long-term forecasts, explainable via advanced visual analytics and accurate, real-time, augmented reality facilities. To achieve its vision, CREXDATA will develop a next generation Prediction-as-a-Service (PaaS) system where action planners will easily register their multimodal data stream sources and compute resource federations and graphically design predictive analytics workflows including (i) data ingestion, fusion, (ii) simulation, (iii) federated learning for pattern extraction and (iv) multiresolution forecasting operators. Decision makers will receive back extremely precise forecasted representations of future worlds reasoned about using transparent AI facilities and with reduced complexity via visual analytics and intuitive augmented reality provided on-site or remotely.
The CREXDATA architecture incorporates several exploitable assets based on cutting edge research, which will significantly outperform the current state of practice in respective fields. CREXDATA will be evaluated in three use cases where real-time critical action planning and timely decision making are of utmost importance: i) maritime domain, for forecasting hazardous situations at sea and impose safer navigational routes, ii) weather emergency management, to allow authorities and first responders proactively act so as to avoid or reduce the impact and speed up recovery from natural disasters, and iii) health crisis management, to limit pandemic outbreaks and come up with non-pharmaceutical means of patient treatment.
The CREXDATA architecture incorporates several exploitable assets based on cutting edge research, which will significantly outperform the current state of practice in respective fields. CREXDATA will be evaluated in three use cases where real-time critical action planning and timely decision making are of utmost importance: i) maritime domain, for forecasting hazardous situations at sea and impose safer navigational routes, ii) weather emergency management, to allow authorities and first responders proactively act so as to avoid or reduce the impact and speed up recovery from natural disasters, and iii) health crisis management, to limit pandemic outbreaks and come up with non-pharmaceutical means of patient treatment.
To facilitate the Complex Event Forecasting, Learning and Analytics over data, we presented an expressive forecasting engine and an (online) optimization method for this engine which can help in determining the proper configuration settings. We presented the simulation scenarios from all three CREXDATA use-cases and a discussion of the methods to be applied for interactively exploring the parameter space of the CREXDATA simulators. We presented a novel bandwidth-efficient technique for Federated Deep Learning and demonstrated its advantage over previous solutions, as well as its applicability in computer vision, which is of crucial importance for the weather emergency use case. We presented an approach which optimizes the execution of arbitrarily many workflows over arbitrarily many devices, under arbitrarily many physical execution options in volatile streaming and network settings. We presented language models developed for monitoring and extracting key information about weather emergencies from social media messages.
To facilitate the explainability, visual analytics and perception of results to end users, we studied and developed explanation models that can be directly applied to black-box models of the use cases of the project. We developed Visual Analytics methods to support the accessibility of the human to the complexity of the models and their uncertainty. We developed Augmented Reality methods, designed to support the interaction of the user with the models and the data, and to provide a more immersive experience in the exploration
of the models and their uncertainty.
We have specified 3 use cases (with several sub-scenarios), worked on simulators regarding the use cases, designed a first version of the CREXDATA architecture and we have made initial progress on integrating some of the CREXDATA tools.
To facilitate the explainability, visual analytics and perception of results to end users, we studied and developed explanation models that can be directly applied to black-box models of the use cases of the project. We developed Visual Analytics methods to support the accessibility of the human to the complexity of the models and their uncertainty. We developed Augmented Reality methods, designed to support the interaction of the user with the models and the data, and to provide a more immersive experience in the exploration
of the models and their uncertainty.
We have specified 3 use cases (with several sub-scenarios), worked on simulators regarding the use cases, designed a first version of the CREXDATA architecture and we have made initial progress on integrating some of the CREXDATA tools.
Compared to the state-of-the-art, we have provided more accurate forecasting models, bandwidth-efficient techniques for federated learning reducing the bandwidth consumption by orders of magnitude, algorithms for optimized data analytics, techniques from extracting information from social media, explanable AI techniques coupled with novel visualization approaches and enhanced information provided to emergency personnel operating on-site.