WP5 - Deciding and Learning How to Act

ICT-48 TAILOR: Foundations of Trustworthy AI integrating Learning, Optimisation and Reasoning

(Working Site)

Welcome

This is the working site of WP5 (Deciding and Learning How to Act) of the ICT-48 TAILOR: Foundations of Trustworthy AI integrating Learning, Optimisation and Reasoning.


WP5 focusses on the fundamental question:

How does an AI agent decide and learn on how to act?

WP5 aims at empowering the agent with the ability of

  • deliberating …

  • how to act in the world …

  • autonomously

WP5 investigates:

  • Reasoning and planning for acting

  • Learning strategies/plans from data

  • Learning models from data, and then do reasoning and planning

  • Learning from past experiences and simulations, for refining strategies/plans or models

  • Monitoring the actual outcome of actions

  • Recognizing possibly unexpected outcomes

  • Reasoning, planning and learning how to deal with unexpected outcomes

  • ...

Crucially, empowering an AI agent with the ability to self-deliberate its own behavior carries significant risks, and this ability must by balanced with safety

The autonomous behavior must be:

  • Guarded by human guided specifications and oversight

  • Verifiable and comprehensible in human terms

  • Ultimately trustworthy (WP3)

Assessing safety becomes essential in this context

  • Formal verification, model checking and automated synthesis fulfilling safety specifications is central to this WP.