Numeric Domain Model Acquisition from Action Traces

Many industrial and commercial applications of planning technology have to reason about numbers. For example, in the area of autonomous systems and robotics, an autonomous robot often has to reason about its position in space, power levels and storage capacities. We call the internal representation that the robot has of its environment its model of the world. It is essential for these models to be easy to construct and ideally, they should be automatically constructed.

This project concerns the subject of learning formal models of state-transition systems from observation of those systems operating. Consider an observer unfamiliar with the game of chess: by observing a sequence of moves, much can be learnt about the rules of the game. Watching a bishop for long enough demonstrates that only diagonal moves are possible for this piece.

Domain model acquisition, or automated modelling, is the problem of allowing a computer to learn its own world model by observing actions. We will develop new methods of automated modelling for state transition systems with numeric state variables. And when we refer to domain model acquisition, we refer to the learning of any state transition system from example data that includes sequences of state transitions, whether that is in the context of automated planning, general game playing, interactive narrative, workplace rostering, or any other type of underlying problem.

We first plan to learn models of domain with a common restriction of numeric variables in planning; the restriction to action costs. This restriction means that each ground action has a constant cost, and that the only numeric variable accumulates the sum of these individual action costs over the length of the plan. This accumulated value is the optimisation variable. Board games with action costs are those in which a score is accumulated throughout the play of the game.

Successful completion of the first stage means that we have a domain model acquisition algorithm to learn models that include action costs. This class of planning domains is an important subset of numeric planning domains. However, many planning domains contain more complex numeric properties and, in particular, arbitrary numeric variables and constraints. The second stage, therefore, will concentrate on developing algorithms to learn these constraints.

Completion of the project will allow models of many different problems to be learnt simply from observation. Examples include such things as capacity limits for certain resources, dimensional constraints for positioning items and strength of friendship level requirement to enable certain actions within a social-network aware interactive narrative setting.

Principal investigator: Dr Peter Gregory
EPSRC First Grant: EP/ N017447/1
Starts: 01 February 2016
Ends: 30 June 2017

More details about the Numeric Domain Model Acquisition from Action Traces project