2023 SR&ED Claim

2023 Scientific Research and Experimental Development tax credit claim

T661 - Part 2 – Project Information

Section B – Project Description

Line 242 -– What scientific or technological uncertainties did you attempt to overcome - uncertainties that could not be removed using standard practice? (Maximum 350 words)

The project is attempting to develop general purpose learning and reasoning software for autonomous robot control. The approach is based on the development of a deep symbolic artificial neural network (ANN). Key requirements include that its decisions must be explainable, grounded on sensors and action devices and that it continuously learns.

Current robot control software is not general purpose. It is specifically developed to function in a predefined environment with predefined tasks to be performed. Also known as narrow AI. An example is Tesla’s self-driving car. Learning capability is limited to driving. General purpose artificial intelligence (AGI) architectures exist but are not designed for robot control i.e., they are not grounded on sense and action devices. Examples include ACT-R, SOAR, CLARION, LIDA, and HTM.  These AGI architectures are also applied only in specifically defined environments. The DARPA Robotics Challenge is a good example of the state-of-the-art robotic control software: https://en.wikipedia.org/wiki/DARPA_Robotics_Challenge.  These architectures make heavy use of stochastic inference which results in behavior that is not explainable.  They also do not use the same structures for pattern recognition and motor action control. Their ability to continuously adapt to changes in their environment is seriously limited.

This research is developing a new cognitive architecture (named Adaptron) to solve these problems. However, there are no simple hierarchical symbolic ANNs based on simple binary nodes that are feed-forward, explainable and use reinforcement learning. There are no ANNs that grow nodes in a hierarchy for recognizing objects, events and for learning and controlling action habits. Even deep learning ANNs, such as convolutional neural networks, do not add nodes dynamically as they learn.

More specifically the project is attempting to develop a new hierarchical symbolic ANN that:

  • Uses binary neurons (binons) with reinforcement learning based on intrinsic motivation.
  • Grows the ANN by adding nodes for parallel (spatial) and sequential (temporal) pattern recognition.
  • Learns continuously - does not separate the training phase from the testing phase.
  • Converts magnitude sensor readings (non-symbolic) into symbolic representations.
  • Learns hierarchies of action habits, performs them in sequence and integrates them with the perception ANN.
  • Has explainable behavior.

Line 244 – What work did you perform in the tax year to overcome the scientific or technological uncertainties described in Line 242? (Summarize the systematic investigation) (Maximum 700 words)

Approach:

  1. Increase the complexity of the ANN structure and algorithms to process the requirements at greater levels of complexity while still handling the lower complexity features.
  2. Run regression tests on already working lower complexity features.
  3. Determine its success at learning and thinking based on the observation of its actions in artificial test environments simulated in software and by inspection of its internal memory traces and processes. More details about the research are available at www.adaptroninc.com.

The following software is being used to perform the research:

  1. About fifty test cases of sequences of symbolic values are being used for event (temporal) pattern recognition using a short-term (working) memory structure and hierarchical long-term memory structure of binons. A particular challenge was to determine when an event pattern had stopped repeating. After some simple techniques were tried it was realized that predicting the next event and determining if it occurred or not was the best approach. Several prediction algorithms were tried but none have yet been able to capture the best prediction based on the historic context of what is in short-term memory and the model of the world represented in long-term memory.

Software and results:         STM-11\STM.vbp,  STM-11\Main.bas, and  STM-11\TestCases.bas.

  1. A new short-term memory experimental test environment was built that allowed for the specification of multi-modal configurations of senses and sensors and for creating test cases based on these configurations. Morse code and handwritten digit test cases have been created. The algorithm for creating and processing the short-term memory has gone through many iterations. The conversion of sub symbolic (numeric) sensory values into symbolic values represented as binons has also been modified.

Documents: STM-Software 3.0 Requirement Document.doc
                     STM Experimental Software - Architecture Document.doc
                     Adaptron Perception - Architecture Document.doc
Pseudocode: Adaptron Perception pseudocode 6th June 2023.doc
Software:      \PycharmProjects\STM-Experimental-Software-Version-3.0
Results:        Adaptron Google drive: Experiments and Results.doc

  1. A new general-purpose robot and grid world with walls and doors and robot simulation environment has been built for testing the best perception algorithms from A) and B) and for testing the action habit learning and control algorithms that have been designed. The world and robot can be configured with multiple sensors (switches, cameras, microphones,  buttons etc.) and action devices (speakers, lights, wheels etc.) such that the robot can interact with the world.

Software:      \PycharmProjects\DAVE-Software
                     DAVE (Deterministic Autonomous Vehicle Environment)
Versions:      Maintained in Github

The robot is controlled by the Adaptron perception-action ANN which runs as an independent program communicating with the world and robot simulation.

Documents:  ONWARD Requirements Docs.doc
                     Architecture Document (ONWARD).doc
Software:      \PycharmProjects\ONWARD
                     (Ongoing NetWork for Adaptron Research and Development)
Versions:      Maintained in Github

  1. We have developed a new theory and design for the learning and control of action habits that integrates with the perception process to continue the execution of action sequences provided the expected outcomes continue to occur. Actions get triggered by the perception process when the current situation is recognized, and a known desirable outcome is predicted. The decision to act results in the activation of all the action habits that are known to produce the outcome but the only one that is executed is the one that is based on the current situation.

Document:    Adaptron and Binons draft book 28th June 2023 - Copy.docx

Line 246 – What scientific or technological advancements did you achieve as a result of the work described in line 244? (Maximum 350 words)

We discovered better temporal prediction algorithms for recognizing sequences of symbolic stimuli. However, none of them have yet been able to capture the best prediction based on the historic context of what is in short-term memory and the model of the world represented in long-term memory.