The AI4PublicPolicy VPME will enable a novel approach to the development of evidence driven policies, characterized by the following properties:

1) AI-oriented: The VPME will take advantage of AI technologies and tools, to boost the automation, efficiency, transparency and trustworthiness of the policy making process. AI4PublicPolicy will integrate a wide range of AI tools, including Policy Recommendation, Policy Simulation and Policy Benchmarking tools, chatbots and NLP tools for the reception of citizens’ feedback, Opinion Mining and Sentiment Analysis tools for analysing citizens opinion and feedback and more. VPME will leverage AI tools across the entire policy development lifecycle. AutoML approaches will also be incorporated in the overall solution to provide the required support to non-technical public administrators while utilizing the respective AI solutions.

2) Cloud-based: Access to resources will be distributed, virtualized and accessible through the EOSC/EGI cloud and HPC resources. EOSC/EGI physical resources will be used to support multiple virtualized instances of the VPME, towards supporting the needs of different users (i.e. policy makers). Hence, end-users of the AI4PublicPolicy services and tools will be able to access the resources that are authorized to use regardless of their owner and location.

3) Local Actors-Centric: The policy development tools and process of the VPME will be centric to the local actors (citizens and businesses), i.e. policies will be driven by the actors’ feedback on the policies under development and on the policy development process itself (e.g. utilization of specific datasets). To this end, the VPME will provide AI tools for collecting and analysing citizens’ feedback, such as chatbot interfaces for citizens interactions and opinion mining tools for analysing and understanding citizens’ feedback in electronic channels like social media. 4) Transparent and Trustworthy: AI4PublicPolicy emphasizes the transparency of the policy making process and of the policy making outcomes (i.e. of the policies). To this end, the project will implement tools that boost the transparency of the AI processes and of the policy outcomes, notably XAI tools that will help policy makers in properly interpreting the outcomes of the AI algorithms and in presenting them to citizens’ accordingly. Coupled with the active participation of citizens’ in the policy development process, the XAI technologies and tools of the project will boost the transparency and accountability of the public authorities, as well as the trust of the citizens to them. To further increase the trustworthiness of the policy development process, AI4PublicPolicy will also implement innovative cyber-defence techniques for the AI tools of the project.

5) Interoperable and Reuse-Driven: The VPME will enable public authorities and other policy makers to reuse policy models and datasets in their policy development tasks. To enable such reuse, AI4PublicPolicy will implement techniques for the semantic interoperability of different policy models and datasets, notably techniques that leverage common ontologies and archetypes, while also realizing the AI models and algorithms as services that can invoked dynamically and thus be reused and repurposed for different datasets.

6) EOSC/EGI-integrated: VPME will benefit from integration with the EOSC/EGI infrastructures. Likewise, the policy development resources of the project will be integrated within the EOSC portal/marketplace and will become available to the EOSC communities. AI4PublicPolicy will also develop its own community that will be offered access to the project’s resources over the EOSC portal.

7) Holistic: The VPME will support a holistic and integrated approach to policy development, through offering not only technical solutions, but also solutions for the ever-important organizational transformation needed to adopt technical (i.e. AI-based) solutions. To this end, the project will produce and integrate (to the VPME) training resources and organizational transformation blueprints that will support the transformation of public administrations in-line with the shift towards data-driven policy making.

8) Ethical and Legal Compliant: AI-based functionalities of the VPME will be compliant to applicable regulations (i.e. GDPR), while adhering to available guidelines for Ethical AI systems. In this direction, the consortium will carry out relevant legal analysis activities, which will be undertaken by the legal expert of the consortium (ALBV).

The AI tools that will be integrated in the VPME and provided to policy makers will include:
1) Policy Development and Recommendation: This tool will enable the execution of AI algorithms based on analytical policy models and datasets that will be integrated in the VPME over an Open Analytics Environment (e.g., KNIME) that act as a data workbench. It will provide the means for defining and executing ML/DL/Reinforcement learning pipelines over policy datasets of the VPME. A variety of AI algorithms will be integrated including deep learning techniques (e.g., ANN (Artificial Neural Networks), Recurrent Neural Networks (RNN), LSTM (Long Short-Term Memory)), traditional machine learning techniques (e.g., decision trees, logistic regression), unsupervised learning techniques for use case with unlabelled data (e.g., K-means clustering) and Reinforcement Learning (e.g., Dynamic Programming, Markov Decision Process (MDP) and Monte Carlo) techniques. The tool will provide the means for reusing algorithms and models across different contexts, by saving configurations at the Open Analytics Environment of the VPME. AutoML techniques will be also exploited to boost the automation and user friendliness of the policy development process for end-users (e.g., policy makers).

2) Policy Simulation and Benchmarking: This tool will enable simulation and (what-if) evaluation of alternative policies on available datasets. It will support the definition and calculation of different KPIs depending on the target policy such as parking space utilization for parking policies, maintenance/repair cost calculation for alternative maintenance policies, revenue generation for alternative urban mobility policies etc. The tool will therefore provide the means for comparing alternative policy parameters and configurations.

3) Opinion Mining, Sentiment Analysis and Document Analysis Tools: These tools will be implemented based on readily available technologies of NOVO and RB. Specifically, NOVO will contribute readily available NLP processing tools over citizens’ comments and surveys that are provided/presented to citizens in the scope of NOVO’s civic engagement platform. RB’s NLP technologies and Text Analytics platform will be used to implement opinion mining and sentiment analysis tools over citizens’ comments on social media, along with document processing tools.

4) Explainable AI (XAI) – Policies Explainability: This tool will boost the interpretability of AI-based policy models, through XAI algorithms that will decompose black-box models (e.g., DL algorithms that are not typically expressed as rules) into rule-based descriptions. Specifically, techniques that explain the operation of deep learning systems based on their dominant features (e.g., Deep Learning Important Features (DeepLIFT) and Prediction Difference Analysis techniques) will be employed and customized to the needs of the policy models of the project.

5) Cyber-defence Tools for Secure and Trusted AI: AI4PublicPolicy will integrate cyber-defence mechanisms that will safeguard the secure operation and the overall trustworthiness of the AI systems. The mechanisms will primarily protect the AI systems against data poisoning and data evasion attacks i.e. they will detect and alleviate attacks against the training data and/or the actual operation of a deep neural network. In terms of data poisoning attacks, the project will leverage XAI techniques in order to identify and unveil malicious neural networks that have been hacked due to the use of fraudulent data for training. XAI will be also used to replay the operations of the network following the identification and removal of fraudulent data sources. Likewise, to alleviate evasion attacks the project will employ: (i) Formal methods that check exhaustively the inputs and output of a deep neural network in order to ensure that it operates appropriately and (ii) Adversarial training techniques, which will use the data that are not properly classified by DL, in order to retrain the network based on new labels to the problematic data (i.e. labels that lead to correct results). The work will leverage readily available toolkits for experimenting with evasion attacks (e.g., AdversariaLib and AlfaSVMLib).