AI Exam Flashcards
AI Paradigms
- GOFAI
- Connectionist
- Behaviour Based
- Embodied AI
GOFAI
Assumes cognition to be computations
Needs Discrete states, unambiguous rules, full knowledge of task and environment
Sense-Think-Act cycle approach
Embodied AI
Balance between an agent neural activity (brain), its morphology (body shape and material properties) and the interaction with the environment. The interaction i influence by the laws of physics (can be limiting but also used to out advantage)
Paradigms of Robot Control Architectures
- Deliberative - Think, then act - finds best course of action but very slow
- Reactive - Don’t think, just act - fast but cannot learn, predict etc, hard to scale too
- Hybrid - Think and act concurrently - fast and dynamic response to changes but comlicated, hard to scale
- Behaviour Based - Think the way you act - hierarchy layered fast and dynamic
Characteristics of behaviour-based systems as an AI
- Situatedness (no prediction, sensorial feedback for change detection bc directly embedded in environment)
- Embodiment (physical setup not theoretical)
- intelligence and emergence ( intelligence comes from interaction with environment, individual behaviours are not intelligent by themselves)
Behaviour Coordination Types
and typical control approaches
- Competitive (Winner takes it all)
- Cooperative (combined action)
Approaches: - Subsumption Architecture (built-in, fixed priorities)
- Dynamic Mechanism (activation levels based on sensor info, goals & other behaviours, highest level has control)
- Motor Schemas (potential field, cooperative method)
Sensing vs Perception
Sensing - The collection of data (detecting a physical object via sensors linke cameras etc)
Perception: The understanding of data (selecting, sensing, organising, interpreting sensations/sensor data)
Control Levels
High Level Control: task planning (get X from the Y), building semantics (in which room is X)
Mid Level: Perception (detection, recognition..)localisation, map building, path planning
Low Level: Sensor reading, actuator control
Breitenberg Vehicles Connection types:
- Ipsilateral (sensors are connected to actuators on the same side)
- Contralateral (sensors are connected to actuators on opposite side)
- Excitatory (sensor reading & signal to actuator are proportional)
- Inhibitory (sensor reading & signal to actuator are inverse proportional)
Sensorimotor integration and Sensorimotor contingencies (SMC)
brain using rapid and selective integration of sensory information to make muscle output
vs
relationship between movements and observed changes in sensory input. Pro active, we predict and make anticipatory movement based on the prediction
SMC Internal Models
Inverse Model: input desired goal, output a plan to achive the goal (anticipatory action). They are limited by feedback delay.
Forward Model: takes a copy of IMs plan as input and predicts what happens when executed. Gives fast response if feedback is delayed
Characteristics of forward models
- state estimation
- giving a prediction before sensor data is available
- correct for errors by transforming them into motor commands
- predict and cancle sensory effects of moving
- cannot only predict input to inverse model but also change the inverse model itself (learning) → adaptive control
Human/Machine Learning Paradigms
- Supervised Learning (agent gets training data incl corresponding target output) -> used for regression and classification (image classification in robotics)
- Reinforcement Learning (learn via trial and error, designer input to correct(rewards&punishment) -> used for video game AI, sensorimotor skill learning (navigation, grasping, manipulation etc in robotics
- Unsupervised Learning (find patterns in data self) -> used for data clustering
What are Neurons, synapses and perceptrons?
Neurons recieve and send electrical signals from and to other neurons.
Synapses are the connections between neurons (it is NOT a physical connection, mircoscopic gap between neurons)
A perceptron is a basic artificial learning unit (artificial neuron)
What is Hebbian Learning?
Correlation between neurons: firing order and time difference. The smaller the time difference the bigger the more the synapses get changed (strengthened if pre synaptic neuron fired first or weakened if post synaptic neuron firing first).
Neuron model: u (input) * w(synapse weight) = v(output)
Synaptic weight change: dw/dt = mu(learning rate) * v x u
NOTE: learning rate is always positive (but less than 1), so right hand side is always positive, so weight only increases
What is differiential Hebbian Learning and what are the two models discussed in class?
To make sure the synaptic change can be negative (like in the biological process) we look at the derivative, bc it can be positive and negative even though the scalar is only positive.
Kosko Model: considers derivative of both in and output, problem: doesn’t consider if the post synaptic node fires first. Symmetric in time
Porr-Wörgötter Model: only derivative of output, not of input, much more realistic (look at graph), assymetric in time
Draw a perceptron
-
Gradient Descent Algorithm
- Find Error = 0.5(t-v)^2
- Find input/weight contribution (how much does the error change when the weight changes. For sigmoid activation dE/dw = v(1-v)(t-v)
- now calculate change in weight like before, but insteat of v use rhs of equation in 2
- learning stops when error is small enough (finding global zero not guaranteed, but local lowest minimal error)
What is AI and what are some AI algorithms/techniques?
A computer solving tasks (thinks&acts) in a way that mimics human behaviour. Main goals: understand biological behaviour, abstract the general principles of intelligent behaviour, apply these behaviours.
Example of techniques:
Decision trees (binary trees), neural networks (human brian structure), clustering, search algorithms like A*
Moravec’s paradox
Reasoning requires very little computation but sensorimotor skills require enormous computational resources
How to design and build behaviour based control architectures?
Behaviour architecture has different abstraction levels, from simple to complex. The behaviours are added incrementally (simple to complex), after every newly implemented behaviour the robot is tested.
Each behaviour can have inputs from sensors and other behaviours, and send outputs to actuators and other behaviours. Behaviours are simple, time-extended processes, not atomic actions.
What is Motion Perception?
Estimation of the speed and direction of elements based on sensory information (which could be visual, vestibular (balance and spatial orientation), propioceptive (self motion and body position)
How are perception and action related?
Intelligent perception can help generate intelligent action and intelligent action can lead to intelligent perception
Breitenberg vehicle drawbacks
only reactive, no predicting and learning, cannot deal with complex situations.
