Skip to content
/ NACE Public
forked from patham9/NACE

Non-Axiomatic Causal Explorer

License

Notifications You must be signed in to change notification settings

andreneco/NACE

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Non-Axiomatic Causal Explorer

Aim

This project builds upon an implementation of Berick Cook's AIRIS, with support for partial observability. The aim is to enhance its capabilities to handle non-deterministic and non-stationary environments, as well as changes external to the agent. This will initially be achieved by incorporating relevant components of Non-Axiomatic Logic (NAL).

Background

Several AI systems, as referenced in related works, employ a form of Cognitive Schematics. These systems learn and use empirically-causal temporal relations, typically in the form of (precondition, operation) => consequence. This approach allows the AI to develop a goal-independent understanding of its environment, primarily derived from correlations with the AI's actions. However, albeit not "necessarily causal" these "hypotheses" are not passively obtained correlations, as they can be re-tested and seeked for by the AI to improve its predictive power. This is a significant advantage over the axiomatic relations proposed by Judea Pearl. Pearl's approach is fundamentally limited, as it cannot learn from correlation alone, but only obtain new probability spaces with a graph of already-given causal relations. This limitation is not present in the cognitive schematic approach, which makes it a more general adaptive learning model better-suited for autonomous agents. Additionally, the use of the NAL frequency and confidence values to represent hypothesis truth value enables efficient revision of the agent's knowledge in realistic settings. Unlike the probabilistic approach, this method can function effectively even with small sample sizes, can handle novel events (unknown unknowns) and has a low computational cost since only local memory updates are necessary.

Architecture

image

Demonstration scenarios

  • Learning to collect salad from scratch: World1
  • Learning how to put the cup on the table, in this case the goal is known to the agent: World2
  • Learning to collect batteries and to pick up keys in order to make it through doors: World3
  • Learning to collect salad with a moving cat as disturbance: World4
  • Learning to play Pong in the grid world: World5
  • Learning to bring eggs to the chicken: World6
  • Learning to play soccer: World7
  • Learning to collect salad while avoiding to get shocked by electric fences World8

Related works:

Autonomous Intelligent Reinforcement Interpreted Symbolism (AIRIS)

OpenNARS for Applications (ONA)

Rational OpenCog Controlled Agent (ROCCA)

About

Non-Axiomatic Causal Explorer

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%