Planning searches for solutions through a general measure of state novelty. When the dynamics of the problem are given through simulator engines such as the Atari games (ALE), GVGAI, or complex robotic and UaV flight simulators, novelty exploration yields state-of-the-art performance compared to known alternatives such as MCTS. Novelty exploration can be combined as well with the exploitation of goal-based heuristics within a general Best First Width Search. BFWS can solve efficiently classical planning problems even when the action model is hidden, opening exciting opportunities to model beyond declarative action representations.

Papers:

(Disclaimer: this is not a comprehensive list, please get in touch to add relevant work that has been missed)

Table of Contents

  1. Classical Planning over STRIPS/PDDL
  2. Classical Planning over Simulators
  3. Planning over Factored Simulators in Functional STRIPS
  4. MDP over Simulators
  5. Multi Agent (Descentralisd & Privacy Preserving) Planning

Classical Planning over PDDL

Classical Planning over Simulators

Planning over Factored Simulators in Functional STRIPS

Factored simulators accept finite domain variables, and mixed Declarative and Programatic (simulated) dynamics

MDP over Simulators

Multi Agent Descentralised and Privacy Preserving Planning

Demo Videos

Simulators screenshots

Atari UaV Robot