2010 SR&ED Claim
2010 Scientific Research and Experimental Development tax credit claim
T661 - Part 2 – Project Information
Section B – Experimental development
240 - What technological advancements were you trying to achieve? (Maximum 350 words)
General purpose learning and thinking software for control of robots based on artificial neural networks (ANN).
- Advancements in the architecture of a hierarchical structure for ANN nodes
- The use of binary neurons (binons) with reinforcement learning,
- Network growth from the addition of nodes for parallel and sequential pattern recognition
- Learnt action habits integrated into an associated hierarchical behavior network
- Conversion of graduated sensor readings into symbolic stimuli
242 - What technological obstacles did you have to overcome to achieve those advancements? (Maximum 350 words)
Current robot control software is not general purpose. It is specifically developed to function in a predefined environment with predefined tasks to be performed. Learning capability, if present, is limited to this environment. General purpose artificial intelligence architectures exist but are not designed for robot control i.e. they are not grounded on sensory input and device activation. Examples include ACT-R, SOAR, and HTM. DRAMA, MAXSON and Brook’s COG are experimental architectures designed for robot control but have not been applied with any success. This research is developing a new architecture to solve the problem. However there are no hierarchical ANNs based on binons that are feed-forward and use reinforcement learning. There are also no ANNs that grow nodes for each new object or sequence recognized or new action habit learnt.
244 - Summarize the work performed in the tax year, and explain how that work contributed to the advancement of scientific knowledge. (Summarize the systematic investigation) (Maximum 700 words)
Approach: Increase the complexity of the ANN structure and algorithms to process the requirements at greater complexity while still handling the lower complexity features. Run regression tests on already working lower complexity features. 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.
Areas of research performed this year - Two pieces of software are being crafted;
A) In the area of object and sequence recognition from multi-sensor senses producing graduated and discrete stimulus readings using a hierarchy of binons.
B) In the area of memory, learning, thinking and action habits.
In area A:
- Evolved the tree structure representation and modified the approach for combining edges and stimulus readings to identify objects, with and without gaps, independent of position, size and intensity. Six versions developed. Discovered it was not possible to maintain a sorted order representation of symbolic stimuli using only the binon tree structure.
- Modified the mechanism for recognizing and representing reflections and negatives of objects.
- Added a way of performing regression tests on pattern recognition testcases, however it became too difficult to keep it up-to-date with changes in the underlying tree structure representation.
- Separated the “Where” from the “What” information in pattern recognition so objects could be recognized independent of size, intensity and position.
In area B:
- Integrated the pattern recognition algorithms from the A software into the B software. This was done with the six different pattern recognition approaches.
- Improved the use of context information in the execution of parallel subconscious sequential pattern recognition. This however was replaced with the new approach in 3. and 4.
- Added a single Short Term Memory (STM) buffer to accumulate and recognize sequential patterns that repeat and then become conscious.
- Replaced STM with one STM buffer per sense and / or independent sensor.
- Changed recognition from sequences of parallel stimuli to parallel combinations of sequences of stimuli.
- Updated the partial matches of trigger stimuli for generalization in learning. Sequential and parallel partial matches were separated.
- Episodic Long Term Memory (LTM) representation was changed from a linear array of experiences built over time to a network structure which maintains the most recent experiences and from which an exact history cannot be recreated.