What is AI? Can you give an example of an AI or AI algorithm/technique?
“It’s the science and engineering of making intelligent machines” - making a computer solve tasks in a way that minic human behavior.
Example AI: in self-driving cars.
Example AI Algorithm: Neural Network
What are the different AI paradigms?
- “Classical” or Good Old-Fashioned AI (GOFAI): Computing with symbols. The hardware doesn’t matter.
- Connectionist: Distributed computational units (neural networks). Loosely resembling the neurons in the brain.
- Behavior based AI: Focused only on the robot
- Embodied AI: Focus on the “system” rather than just the robot
What is GOFAI? Can you give an example of a GOFAI algorithm/technique?
GOFAI: “Good Old-Fashioned AI”:
Assumes:
- Cognition is a computation: intelligence is an algorithm.
- Computing with symbols (in chess a symbol could be a piece, its placement and a set of moves).
Example: Heuristic search with the A* algorithm.
What kind of problems can we use GOFAI for?
Any problem that requires logic: if the problem can be described with symbols, it can be solved with GOFAI.
If a solution exists, it can be found by any search algorithm such as A*.
Example: Chess:
- Discrete states: Set of locations of all pieces on the board at any given time
- Unambigious rules of transitions from one state to the next: The pieces can only move in specific, pre-determined ways.
- Full knowledge (model) of the task and environment: Task is to win (capture the opponent’s King). Environment is the board and all the pieces.
What are the drawbacks of GOFAI?
GOFAI is NOT human intelligence:
- It tries to solve problems in a fixed world with a fixed set of rules.
- The world is unpredictable => must be able to adapt.
- The world is a dynamic system, can’t be defined by a set of rules.
Moravec’s Paradox (explanation): “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility”
What is Moravec’s paradox?
Reasoning requires very little computation (e.g., playing chess), but sensorimotor skills require enormous computational rescources.
What kind of problems are difficult for GOFAI? Why is it difficult? Can you give an example of such a problem?
- Real world problems that require sensory motor skills, or a combination of more smaller problems
- Continuous problems:
□ The world is continuous
□ Cannot understand the world as the dynamic system it is and respond to it.
Example: Tell a computer to take the train. - Navigate to the train:
- Get to the right platform
- Get to the right place
- Difficult to navigate with all the flashing light and all the people.
How is GOFAI typically applied in robotics? Can you give an example?
- Separate the low-level controller from all the intelligent part.
Sense-Think-Act repeated in sequence: - Sense: Collect sensor data.
- Think: Make a decision based on sensor data from the environment.
- Act: Carry out the decision
What is embodiment? Can you give an example of embodiment in nature/machines?
- A body is required for intelligence.
- Embodiment focusses on the interaction between “physical processes” and “information processes”
- Embodied systems are subject to the laws of physics
Enbodiment in Machines: Interaction between robot’s actions and its control program. (Drone flying in a windy environment. Can exploit the wind or have to fight it).
Embodiment in Nature: Interaction between physical actions and neural processing, or between the body and the brain (Arctic Fox: can’t see rodents => move head to hear rodent. Located rodent => jump at the right time to catch it)
How is embodiment useful?
Designing the body to exploit the laws of physics can simplyfy the controller design.
- Tasks can become easier when embodiment is considered
- Proper sensor placement: Place sensors where it is needed the most => Computaitonal requirenments are reduced by exploiting the physics of movement relative to the sensory input
- Exploit material properties: Offload as much of the computation in the body of the robot to lessen the computational power needed in the controller.
What is embodied AI? Can you give an example of embodiment AI in machines?
- The interplay between an agent’s neural activity (the brain), its morphology (the body’s shape and its material properties), and its interaction with the environment, and that interplay can be exploited to achieve certain tasks.
- The laws of physics in the environment influence the interaction (the dynamics of perception, control, movement, learning). Can be limiting, but also exploited.
- The body morphology (shape, structure and materials) influence the interaction: designing the body to exploit the laws of physics can simplyfy the controller design.
It is a system: Three Layers: Brain, body, environment.
Example:
- D-shaped Roomba = compromise between cleaning the corners of a room and avoid getting stuck.
- Round objects does not get stuck as easily => controller does not neeed to deal with the robot getting stuck).
What are the 3 layers in embodied AI?
- Brain: Information self-structuring (controller)
- Body: Body dynamics / morphology (Mechanical system & Sensory system).
- Environment: Task-environment
How does the environment contribute to intelligence in embodied AI?
The laws of physics in the environment influence the interaction between the brain, the body and the environment (the dynamics of perception, control, movement, learning)
Intelligent behaviour arises from this interaction.
Have to consider all components in the system (physical processes and informational processes, not just the algorithms)
=> Focus on interaction/coupling between the brain, body and environment
=> Only way to generate intelligent behavior
How is embodied AI different from GOFAI?
GOFAI: No body. The algorithm is everything.
Embodied AI: Intelligence requires a body. The whole system (brain, body, environment) creates the intelligent behaviour.
