Comirit: Commonsense Reasoning by Integrating Simulation and Logic Benjamin Johnston and Mary-Anne...

Post on 02-Jan-2016

225 views 1 download

Transcript of Comirit: Commonsense Reasoning by Integrating Simulation and Logic Benjamin Johnston and Mary-Anne...

Comirit: Commonsense Reasoning by Integrating Simulation and LogicBenjamin Johnston and Mary-Anne Williams

Commonsense Intelligence and Reasoning through Integrative Technologies

Motivation

> Product: Middleware/infrastructure> Function: Real-world Commonsense

> Robot house-keeper> Robot rescue> [later...] Symbolic services for business software

> Inspiration: Failure of logical commonsenseHolds(Contains(oc,o),t) &&

Holds(Downward(oc,t)) &&

(Holds(Liquid(o),t) ||

Holds(smaller(shape(o), VirtualLid(shape(oc))),t))

=> Happens(Fall(o,oc),t)

Philosophy

> The technology already exists[but,]> It just needs clever exploitation[how?]> We do this by thoughtful integration

> Careful selection> Careful reinterpretation

My Proposal: Comirit

Graph Based

Simulation

Tableaux-based Logical

Deduction

Imagination,Iconic Reasoning

Symbolic Reason,Abstract thought

Graph-based Simulation

> (Hyper-)Graphs are very general> Can represent any structure from 3D

wireframe models to social networks and document structures

> Each node and edge is annotated> Simulation is iterative update on annotations> 3D models: mass and springs

Graph-based Simulation

Graph-based Simulation

Graph-based Simulation

Graph-based Simulation

Graph-based Simulation

Graph-based Simulation

Graph-based Simulation

Graph-based Simulation

Graphical Simulation: Advantages

> Computation is local> Few laws can capture many domains> Laws are largely orthogonal> Laws are simple to implement

Logical Deduction with Tableaux

Logic provides:> High level control> Abstract thoughtAnalytic tableaux:> Efficient theorem proving on typical problems> Great success in Semantic Web reasoning> Syntactic search for (counter-) models of a

formula> Regular mechanism of tree search

Analytic Tableaux

Coffee and (Slow_Move or Fast_Move) and not(Mess)

Analytic Tableaux

Coffee and (Slow_Move or Fast_Move)not(Mess) Mess

Analytic Tableaux

(Slow_Move or Fast_Move)not(Mess)

Coffee

Messnot(Mess)not(Coffee)

Analytic Tableaux

Slow_Movenot(Mess)

Coffee

Messnot(Mess)not(Coffee)

Fast_Movenot(Mess)

Coffee

Key Insight

> Simulations deduces consequences of a (narrow space of) scenarios or possible worlds

> Tableaux systems search over the space of scenarios or possible worlds

[gives us an insight!]> Simulation (and other reasoning) can be made

an operation in a tableaux system

Integration

Implementation

Simplicity, cohesion, open-endedness:> Everything goes into the tableaux system> Even the process of tableaux reasoning!

> (Anything you can do, I can do meta)

Future Work

> Integration with other methods> Heuristics> Logics> Non-logical reasoning> Self organization

> Online learning of simulation> Cognitive architecture