Comparison

From TinyCog
Jump to: navigation, search

How does Scene Based Reasoning compare with other Cognitive Architectures? In this chapter we reproduce a number of classification approaches from literature and specify the approach of SBR.

Comparison by Application Domain

Cognitive architectures are software systems built to work in a number of specific use-cases or application domains:

  • Modelling (aspects of) the human mind
  • Robotics and object handling
  • Logistics and system's coordination
  • Natural Language Processing (NLP)
  • Navigation and path-finding
  • Playing video games

Scene Based Reasoning currently focuses on modelling aspects of the human mind and playing video games. The next step is to incorporate NLP.

Comparison by Approach Towards Consciousness

[Reggia 2013] proposes several approaches to model consciousness:

  • Global Workspace
  • Information integration
  • Self-models
  • Higher-level representations
  • Attention mechanisms

Scene Based Reasoning assumes the self-model approach, combined with an attention mechanism that allows the system to focus on it's self-models.

Comparison by Reasoning Architecture

[Duch, Oentaryo, Pasquier 2008] proposes a number of architectures that are used to build cognitive architectures:

  • Production System
  • Semantic Network and Graph Database
  • Neuronal Network
  • Other statistical/stochastic reasoning
  • Other reasoning

Scene Based Reasoning uses a number of different reasoning mechanisms, even though only the DL system is implemented as of today:

  • Description Logic (semantic network) is used for "1st Mind" reasoning:
    DL is used to model the relationship between objects including is-a class hierarchies, is-part-of object composition and other properties and relationships.
  • Statistical [Factor Analysis] for plan learning:
    Historic plan execution performance is analyzed to extract the parameters favoring successful execution. Example in the Hunter Domain: The success of bow-and-arrow hunting depends on the distance to the object and the time of the day (less successful during darkness).
  • Production Rules for "2nd Mind" reasoning:
    Together with the Attention Subsystem, production rules are learned as shortcuts for time consuming planning processes. Example in the "Formula Simplification" domain: Rules to simplify arithmetic expressions.

Comparison by Environment

[CogArch.org] defines a number of types of environments in which a Cognitive Architecture may act:

  • Static Environments vs. Dynamic Environments
    Refer to the importance of planning for successful task execution. Very dynamic environment may even include changing object ontologies.
  • Simulated Environments vs. Real-World Environments
    Agents in real-world environments need to deal with ambiguities and noisy sensor data.
  • Complex Environments with rich knowledge or input

TinyCog is specifically designed for dynamic environments thanks to it's strong focus on planning and Plan Recognition Subsystem. The authors believe that the 3D Reconstruction subsystem will provide the base for operation in real-world environments, even though this claim needs still to be verified. Environmental events can be handled despite being asynchronous and unpredictable, with the case of concurrent events not yet determined. Also, there are no results yet for knowledge- or input rich environments. TinyCog is currently conceived to execute only one task in parallel.

Similarities with Other Architectures

Scene Based Reasoning clearly classifies in the family of cognitive architectures that began with SOAR and ACT-R. Particular similarities exist with the [ICARUS] system [Langley et. al. 1991] with respect to the explicit representation of plans with decompositions, the "grounding in physical states", the purpose of controlling a physical agent, the use of observable attributes as a semantic base and spatial roles/ relationships between objects. Both systems share a development roadmap that includes modeling social interaction [Langley 2013].

Special Focus of TinyCog

TinyCog can be seen basically as a collection of some 40 "best of breed" formalisms from almost every AI discipline that are adapted to work together on the base of Scenes and Plans. TinyCog does not pretend to be based on "the one and only mechanism" that explains them all. Instead it focuses on the common data-structures and semantics necessary for the integration of a large number of subsystems into one working system. The selection of subsystems follows the authors' understanding of the "functional requirements" of an AGI and an "everything is a plan" intuition about the structure of the human mind.

Everything is stored as a Scene

Scenes are used in TinyCog as a data-structure for representing almost everything including sensor data, mental images, semantic networks, Description Logic objects and Planner states. The PR guys will be happy if we tell them that "TinyCog can dream in images". It will also simplify the visualization and debugging of Plans because the execution path can be converted into a movie clip.

Everything is a Plan

Using a Plan to encapsulate both perception and action, most subsystems become self-evident:

All other Subsystems perform auxiliary functions supporting these main functions.

Progress in Planner Research

Scene Based Reasoning does not specify any specific Planner. In fact it is possible to use specific planners for certain types of problems, resembling specialized neuronal circuitry in the brain. It also allows advantage to be taken of task decompositions and other recent developments.

ToDo: Recent developments