Ruleml2012 - A production rule-based framework for causal and epistemic reasoning

A Production Rule-Based Framework for Causal and Epistemic Reasoning

Theodore Patkos, Abdelghani Chibani, Dimitris Plexousakis, Yacine Amirat ({patkos, chibani, amirat}@u-pec.fr, [email protected])

Laboratoire Images Signaux et Systèmes Intelligents (LISSI) University of Paris-Est Creteil, Paris

Institute of Computer Science –

Foundation for Research and Technology Hellas (FO.R.T.H.)

6th International Symposium on Rules (RuleML’12),

Research Work supported by the ITEA 2 EU Projects: A2NETS, PREDYKOT

Outline

• Overview

• (Epistemic) Action Theories

• Production Rule-based Framework

• Application Domain

• Conclusions

RuleML 12 2

3

Motivation and Objectives

• Action theories and production systems have widely been used in KR&R to represent knowledge change in dynamic domains.

• Our objective is to exploit the expressive capacity of logic-based theories and the efficiency of rule-based systems, reconciling their differences.

• The outcome is a complete framework that performs runtime reasoning about events, knowledge and time in expressive domains.

• The system can carry out temporal projection and deductive narrative verification tasks and is evaluated in real-world settings

RuleML 12 3

4

Framework overview

Event Calculus • Reasoning about action and time • Solution to problems (frame,

ramification, qualification) • Commonsense phenomena

DECKT • Epistemic reasoning • Hidden causal dependencies, rather

than possible worlds structures • Sensing, potential actions etc

Rule-based forward-chaining production system

• NaF, semi-destructive update • Salience values, subsumption…

• Online/offline reasoning • Multiple model

generation • GUI/Java interface

Application Domain • Ambient Intelligence, AAL • Benchmark problems (e.g.,

Shanahan’s circuit)

Th

eore

tica

l fo

un

dat

ion

s

Imp

lem

enta

tio

n

Co

ntr

ibu

tio

n

RuleML 12 4

Outline

• Overview

• (Epistemic) Action Theories

• Production Rule-based Framework

• Application Domain

• Conclusions

RuleML 12 5

6

Background – Action Theories

• The objective is to express the dynamics of the world

• Therefore always include a more or less implicit general notions of time,

change and causality.

• Action Theories automate the

process of commonsense reasoning, in order to • predict the outcome of a given

action sequence • explain observations • find a situation in which certain

goal conditions are met.

RuleML 12 6

7

Background – Commonsense phenomena

• Related issues • Representation • Effects of Events • Indirect Effects of Events (Ramification problem) • Context-dependent Effects • Non-deterministic Effects • Concurrent Events • Preconditions • Inertia (Frame problem) • Actions with duration • Delayed Effects and Continuous Change • Default Reasoning (Qualification problem) • …

RuleML 12 7

Background – Discrete Time Event Calculus

• The EC distinguishes three kinds of objects – events, fluents and timepoints. To do this, it appeals to a sorted first-order language.

• Commonsense Law of Inertia: things tend to persist unless affected by some event.

Positive and Negative Effect Axioms (Σ)

fi C [HoldsAt(fi,t)] Initiates(e,f,t)

fi C [HoldsAt(fi,t)] Terminates(e,f,t)

State Constraints (Ψ) i [HoldsAt(fi,t)] HoldsAt(f,t)

EC Predicates

• HoldsAt(f,t)

• ReleasedAt(f,t)

• Happens(e,t)

• Initiates(e,f,t)

• Terminates(e,f,t)

• Releases(e,f,t)

RuleML 12 8

Epistemic Action Theories

• Epistemic (modal) logic: An agent is

said to know a fact if this is true in all possible worlds.

RuleML 12 9

Epistemic and Causal Reasoning

• Action theories that do not model time explicitly have been extended to reason about knowledge in a straightforward manner. • The Situation Calculus

[Moore 1985, Scherl&Levesque 2003]

• The Fluent Calculus [Thielscher 2000]

• The Action Language Ak [Lobo et al. 2001]

• For example, suppose an action E that makes F true if F’ is true:

RuleML 12 10

Theodore Patkos et al. 11

DECKT – Hidden Causal Dependencies

• We developed a unified formal theory for epistemic, causal and temporal reasoning

• able to express diverse phenomena of commonsense reasoning, about knowledge

• still being computationally feasible.

E

RuleML 12 11

Outline

• Overview

• (Epistemic) Action Theories

• Production Rule-based Framework

• Application Domain

• Conclusions

RuleML 12 12

13

EC Jess-based Reasoner

RuleML 12 13

14

Requirements and Challenges

• There exist different implementations of the Event Calculus for offline reasoning, • but certain of its features are not appropriate for runtime execution

(e.g., explicit frame axioms – computational frame problem)

• DECKT has been designed for practical implementations, • Therefore introduces added features not typically met in Event Calculus

(e.g., reification of formulas, time-dependent meta axioms for handling of HCDs)

RuleML 12 14

15

DECKT Time-Dependent Meta-Axioms

fi C [HoldsAt(fi,t)] Initiates(e,f,t)

• Event e initiates f if f’ is true • (KT6.1.1) handles the case of

unknown preconditions

• Circumscription at every timepoint would be inefficient.

e

RuleML 12 15

16

Production rules with high level structures

• Features that we employ with rule-based reasoning involve • Dynamic rule construction to accommodate HCDs • Semi-destructive update of the KB, where applicable • NaF, rather than circumscription • List handling • Salience values for conflict resolution (e.g., concurrent events). • Numerical manipulations

RuleML 12 16

17

• Released fluents and NaF

Operational semantics – Model generator

Sensing and

Acting

KB

DEC/DECKT axioms (destructive update)

Ramifications (state constraints)

Triggered events

Alternative Models

RuleML 12 17

18

Non-Epistemic Reasoning Cycle

KB of Fluents

Released

Released

Released Constrained

Effect axioms

fi C[HoldsAt(fi,t)] Initiates(e,f,t) fi C [HoldsAt(fi,t)] Terminates(e,f,t)

State Constrains

i [HoldsAt(fi,t)] HoldsAt(f,t)

.

.

.

Trigger axioms

.

.

.

.

.

.

fi C [HoldsAt(f1,t)]

Happens(e,t)

Sensing

X

X X

RuleML 12 18

T=t T=t+1

19

Epistemic Reasoning Cycle

KB of Epistemic

Fluents

Effect axioms

State Constrains

Trigger axioms

Sensing

RuleML 12 19

T=t T=t+1

20

An Informal Note on Complexity

• The logical theories set a lower bound as to how efficient an

implementation can be.

• The predominant computational complexity factor… • for the non-epistemic case is the number of released fluents • for the epistemic case is the set of HCDs

• Query answering on ground facts is of linear complexity.

• Jess pattern matching depends on the syntactic form of rules. • Ranges from O(p) to O(pn)

RuleML 12 20

21

Framework overview

Event Calculus • Reasoning about action and time • Solution to problems (frame,

ramification, qualification) • Commonsense phenomena

DECKT • Epistemic reasoning • Hidden causal dependencies, rather

than possible worlds structures • Sensing, potential actions etc

Rule-based forward-chaining production system

• NaF, semi-destructive update • Salience values, subsumption…

• Online/offline reasoning • Multiple model

generation • GUI/Java interface

Application Domain • Ambient Intelligence, AAL • Benchmark problems (e.g.,

Shanahan’s circuit)

Th

eore

tica

l fo

un

dat

ion

s

Imp

lem

enta

tio

n

Co

ntr

ibu

tio

n

RuleML 12 21

Outline

• Overview

• (Epistemic) Action Theories

• Production Rule-based Framework

• Application Domain

• Conclusions

RuleML 12 22

23

Application Domain – Ambient Intelligence and Ubiquitous Robotics

• Sensor-rich collaborative environments

• Temporal constraints are ubiquitous

RuleML 12 23

24

Conceptual Layers – Knowledge representation in a smart space

• Moving from low-level data to high-level knowledge inference more expressive tools are needed

• AI has a decisive role to play: • representation of

contextual knowledge, • context inference,

• collaboration of devices to achieve common objectives,

• planning in dynamic domains,

• commonsense reasoning

RuleML 12 24

25

High-level Activity Recognition – Challenges for the State-of-the-Art

• Semantic Web tools (ontologies and rule-based reasoning) are widely used

to tackle AmI-related problems

• Complex ambient systems test the limits of these methods in terms of expressiveness and reasoning capacity.

• Agents inhabiting smart spaces need to exhibit • Cognitive skills and commonsense reasoning • Temporal reasoning • Operate under partial observability

RuleML 12 25

26

Reactive Reasoning and Temporal Projection

RuleML 12 26

Outline

• Overview

• (Epistemic) Action Theories

• Production Rule-based Framework

• Application Domain

• Conclusions

RuleML 12 27

28

Contributions and Ongoing Work

• The Event Calculus provides a declarative specification of state transitions,

while DECKT provides a Kripke-equivalent epistemic semantics

• Production rules obtain high-level structures.

• We aim at a tool for both educational and practical use.

• We extend the editor to support the full expressive power of the EC

• Benchmark and use case evaluation of real settings is our ongoing work, considering further enhancements of the system

• We study its integration with probabilistic methods of inference.

RuleML 12 28

The end

Thank you for your attention!