Animal learning

Norbert Weiner (author of `Cybernetics') identified two phenomena that characteristic living systems:

• the power to learn.
• the power to reproduce.

Why is artificial intelligence so concerned with chess machines when Elephants don't play chess? (Title of a paper by Rodney Brooks[brooks1990elephants])

What is intelligence?

• Is animal intelligence different to machine intelligence?
• Who decides what is intelligence?
• What shades of intelligence?

Howard Gardiner identified 8 types of intelligence

• Linguistic
• Logical/Mathematical
• Spatial
• Bodily-Kinesthetic
• Musical
• Interpersonal
• Intrapersonal
• Naturalist.

But there can be other taxonomies of intelligence types, e.g. for machines

• ability to play chess
• ability to Infer answers. Turing test, chatbots, 20 Questions games, (http://20q.net/)
• ability to recognise patterns (face recognition software)
• robotic (ability to use tools)
• Language understanding and translation - e.g. google translate

Does the brain give us the ability to predict outcomes?

The answer is probably yes. If so the key question is

• How does the brain give us the ability to predict outcomes?
• See Miall[Miall93:_is_cereb_smith_predic]

Resilient machines (2006)

Key paper is Bongard, Zykov and Lipson 2006[bongard2006resilient] other papers are lipson [bongard2006automated]

The robot is the starfish. It has a fixed morphology consisting of

• 8 joint sensors,
• 2 tilt sensors and
• 8 joint actuators.

• Most robotic systems get the humans to construct a mathematical model of the robot kinematics and/or dynamics and then plan movement against that model.
• Approach is expensive (time to construct a valid model), and requires calibration
• Robot does not have an explicit model of itself
  • Methods like SLAM model and adapt to change the environment
  • Very few robots allow adapt to changes to their own morphology
• Premise, create multiple internal models and use a system identification like approach to select the best
• Don't freeze the system identification process during use
  • Rather look for disagreements between model and sensors (this is how Kalman filters work!)
  • When disagreement is high, re-initiate the 'model generation and evaluation' cycles.

(https://www.creativemachineslab.com/evolutionary-self-modeling.html)

supplemental material

Video

The 'self modelling video has the following parts':

• Part I three cycles of model synthesis and action synthesis (Fig. 1A-C).
• Part II locomotion synthesis using the best self-model (Fig. 1D).
• Part III the physical robot executing the best behaviour (Fig. 1E).
• Part IV a sample experiment after the robot suffers damage. The robot is shown alternating self-modelling with exploratory action (Fig. 1A-C); then, the best compensatory gait is shown running on the self- model (Fig. 1D), after which it is executed by the physical robot (Fig. 1E).
• Part V some other compensatory gaits,