This page is meant to be an accessible entry point to what is meant by “Integrative AI”, and to the resources on Integrative AI that are available in the AI4EU AI on-demand platform. This guide is part of the broader AI4EU scientific vision on “Human-centered AI”, available here.

What is “Integrative AI”

Consider the problem of planning maintenance activities for a fleet of industrial machines: this task requires paying attention to minute signs of stressed components, so as to best schedule repair operations. Predictive approaches based on Deep Learning and vast amounts of data can be very accurate for this kind of problems, at the cost of being opaque. However, there is a different possible approach: mathematical formulas that describe the system as precisely as possible, using statistics and decades of empirical knowledge collected in the field.

As in this example, it often seems inevitable to make a choice between more understandable and controllable methods, and the power of learning in AI. But what if this were not strictly necessary? What if we could have accurate _and_ understandable predictions? What if we could take advantage of _both_ huge amounts of data _and_ human designed models and rules?

The goal of Integrative AI is to combine AI methods to improve beyond their individual strengths and to compensate for their individual weaknesses.

Here are a few problems arising in human-centered AI that are likely to benefit from an integration of symbolic and sub-symbolic AI approaches:

  • Consider a neural network that controls traffic lights. We might want to ensure that this system never gives two crossing lanes green light at the same time, and that each lane eventually gets green light in a given time span. An integrative approach in this case would enable using (e.g) Reinforcement Learning to provide adaptability, while safety and fairness constraints can be handled via First Order Logic, or Answer Set Programming, or Constraint Programming.
  • Consider navigating a maze that has obstacles cluttering the floor while carrying a stack of items that is sensitive to sudden movements. Tackling this problem requires a combination of spatial and temporal reasoning, from the low level of individual motions to the high level of devising an abstract plan, and rather different techniques are best suited to deal with the different levels.
  • Consider a system based on Deep Learning that matches job opportunities and potential candidates. A user of the system will care about matching quality (including its fairness), but also about the motivations for its behavior. Explaining decisions may become easier if expert-designed knowledge can be injected into (say) a Deep Network, or if a Deep Network can be manipulated via symbolic reasoning techniques.
  • Consider an epidemic control problem, where we need to choose how to best contain the spread of some disease. Simulation models are an effective tool to assess the efficacy of containment measures, but evaluating all possible combinations of measures takes prohibitively long. We could however use machine learning (ML) to approximate the behavior of the simulator, and then use optimization methods (e.g. Mixed Integer Linear Programming) to  quickly find the best combination.

In general, Integrative AI consists of approaches that combine AI systems based on heterogeneous formalisms and algorithms.

Resources on Integrative AI available on this platform

Below is a list of useful resources on Integrative AI currently available on the AI4EU platform. The list is by no means complete, but it can be a good starting point: any contribution or suggestion for further content is well accepted!

Background knowledge on Integrative AI

  • K1: We maintain an extensive survey on the current research in the field of Integrative AI. This is a living document, updated by the researchers in AI4EU every six months.
    • Link to the current version:  [coming soon]

Tools for Integrative AI

  • T1: EML (Empirical Model Learning) translates a trained machine learning model into combinatorial optimization formalisms. This enables reasoning about the trained machine learning model in combinatorial optimization. It could be used to handle the epidemic control examples, to check safety and fairness and to generate counterexamples (e.g. traffic control or job matching).  The AI4EU platform hosts the MELib python library for EML, together with pointers to papers, a tutorial, and source code.
  • T2: Hexlite (Answer Set Programming with External Computations) is a combinatorial optimization solver based on Answer Set Programming that permits to integrate arbitrary Python code as external truth computations, thus enabling heterogeneous reasoning. It could be used to handle our maze navigation and the epidemic control examples. The Hexlite solver is available on the AI4EU platform, together with links to additional resources and to the HEX manual. It is implemented in Python and provides a Python and a Java interface for external computations.
  • T3: SBR (Semantic Based Regularization) integrates ground logical formulas into the cost function of a Kernel Machine learning model. As a result, the model learns both from given data points and from the structure provided by the expert. It could be used to ensure fairness and safety in the traffic control and job matching use cases, and to integrate expert knowledge in predictive maintenance. A docker container for examples of how to use SBR is available on the AI4EU platform, together with pointers to additional resources.
  • T4: Meta-CSP: This AI framework provides infrastructure for combining multiple Constraint Satisfaction Problems (CSPs) into a single CSP so that the problem is solved hierarchically, potentially with solvers specialized for different problem classes. It could be used to enable multi-level reasoning in the maze navigation use case. The Meta-CSP framework is available as a Java library on the AI4EU platform.
    Link to the tool:
    Uploaded by: Federico Pecora [R7]
  • T5: Coordination_ORU: This AI tool integrates motion planning techniques with high level conflict resolution among multiple robots. It permits online addition and modification of goals for individual robots and ensures that all robots are controlled in a way that they will not collide with other robots, no deadlocks will occur, and all goals will be achieved. It could be used to handle scenarios similar to the maze navigation problem, but that involve multiple agents/robots. The Coordination-ORU tool is available on the AI4EU platform as Java code designed  for ROS (Robot Operating System) integration.
  • T6KENN (Knowledge Enhanced Neural Networks) is a library for python 2.7 built on top of TensorFlow that permits to enhance neural networks models with logical constraints (clauses). It does so by adding a new final layer, called Knowledge Enhancer (KE), to the existing neural network. The KE changes the original predictions of the standard neural network enforcing the satisfaction of the constraints. Additionally, it contains clause weights, learnable parameters which represent the strength of the Constraints.

Data sets for Integrative AI

Coming soon.

Case studies of Integrative AI

Coming soon.

Groups related to Integrative AI

Researchers working on Integrative AI


Please help us to complete and maintain this document by notifying corrections or addition to the document maintainer, Michele Lombardi.

Note: if you want to add a software resource, data set or researcher to this document, you first need to make sure that they are available in the AI4EU platform, e.g., by publishing the software.


Cite this document

This document is published under the Creative Commons License Attribution 4.0 International (CC BY 4.0).  It should be cited as:

  • Michele Lombardi (editor), “A simple guide to Integrative AI”. Published on the AI4EU platform: June 22, 2020.