Presentation is loading. Please wait.

Presentation is loading. Please wait.

Introduction to Artificial Intelligence Revision Session.

Similar presentations


Presentation on theme: "Introduction to Artificial Intelligence Revision Session."— Presentation transcript:

1 Introduction to Artificial Intelligence Revision Session

2 This course is about practical AI Started off by looking at ideas of what AI could be, attempts to reproduce consciousness, strong vs weak – Turing test – Searle’s Chinese room argument – Experience with chatbots We’re clearly years (if ever) from achieving human like intelligence

3 Problem Solving So this course is about: how people solve problems and do tasks How we can get machines to do it for us – Means some formalisation: ideas=> algorithms In term 2 we also looked at how nature does some problem solving

4 How do people solve problems? I used lots of examples in tutorials e.g. driving cars, planning routes, table tennis People solve big problems by breaking them down into little ones Different types of problems use different mental tools (different parts of the brain)

5 5 Computers as black boxes Model InputOutput A way of grouping problems according to missing element: to give us extra insight so we can identify appropriate technology

6 Example: road sign recognition Experienced driver does it subconsciously, visual inputs trigger stimulus of concept Given suitable cameras &image processing, an AI system could do this with: – a trained Artificial Neural network – A set of rules hard coded by an expert – A decision tree – Finding the “nearest neighbour” from a set of examples

7 Learning to recognise road signs Humans do this quite quickly, – because they are designed to exploit earlier learning e.g. Recognising numbers Machines can do this too: – We provide them with lots of labelled examples and they learn to build/refine a model: training the ANN with back-propagation, inducing the rule set, Constructing or evolving a decision tree

8 We can even have an optimisation component Humans quickly learn what to look for in the images and what bits to ignore to help us generalise – As babies we learn to focus our eyes... There is an optimisation task: – selecting and tuning the image processing that goes into the model. – whether that model is learned or provided by humans

9 Optimisation / planning problems Given a model of the system /world (or the logic that governs it) task it to find the (series of) inputs that meets some desired goal Route planning, robot/car actuators, planning (e.g. Cannibals/hostages), shot selection, deciding on a game strategy Tree search (depth-, breadth..), hill-climbing, Genetic algorithms

10 Model –building problems Given a set of experiences (stimulus- response, cause-effect,...) task is to find a model that maps inputs onto the right outputs Maze learning for re-use, learning to recognise things, characterising opponents, diagnostic support,... Rule induction, back-prop to train an ANN, using Genetic Programming to evolve trees

11 Simulation/classification problems Given a model of how the world works (rule base, ANN, first order logic, “rules of game”) and some set of inputs (measurements of state of world) find the appropriate output Sudoko, recognition, decision support, logic problems, applying game strategy Feed inputs into trained decision tree/ANN, forward chaining rules in expert system

12 Common theme: search to find right set of moves to take best weights for a given ANN structure right sequence of rules to apply Can be considered as finding the right one from a whole set of alternatives

13 How do people do this? Lots of evidence that early learning (recognition, control of arms etc.) is done by trial and error with corrective signals Many “conscious” activities are constructive: – Series of decisions considering alternatives – With backtracking Not all: e.g. route finding from map, “brainstorming”

14 How can we programme this into a computer? Term 1: We looked at approaches based on how we think people do things logically – State-based search using finite State Machines – Logic programming, depth-first, breadth-first, A* – typically thought of as constructive Term 2: Nature inspired approaches – Iteratively refine “whole” solutions – apply set of rules to move between/ generate solutions based on some fedback (quality)

15 Example: “blind” depth-first search Humans: series of decisions Computer: series of assigning values to variables Humans: explore one solution and backtrack if necessary – Exploring a maze, sudoko with a rubber Computer: depth-first search using stack to store nodes (decisions made)

16 Example – maze

17 Alternative state search methods Breadth-first if no information to guide search Hill-climbing, if there is information to help make a decision about which route to take out of node best-first, A* as more complex (memory) variations

18 Knowledge Representation Sometimes use Finite State Machines Sometimes useful to explicitly store and manipulate knowledge, facts – Knowledge representation – Semantic nets, frames, semantic web

19 Artificial Neural Networks The human brain starts off with a model of the world and then improves it. ANN’s based on perceptron as highly stylised model of a single neuron Configuration + weights = model Back-prop as a way of changing weights (moving between models) to minimise error on training set

20 Evoltionary Computation Generic search mechanism based on evolution Species encoded as DNA= solution encoded in binary/vector/lisp Typically used for optimisation, but not always – E.g. Could evolve weights for ANN – GP evolves trees that could be programmes, decision trees

21 General Scheme of EAs

22 Swarm Computing Means to automate the creation of useful behaviours via lots of really simple interacting things. – Search behaviour e.g. Ant colonies, evolution, particle swarm – Can get lifelike behaviours with minimal programming – good for animations etc How do you find the right rules? – Expert system, evolution?

23 AI in practice Break your problem down into little bits Characterise each sub-task – What type of problem is it, – What type of technology might be appropriate – How might you encode a (partial) solution? Don’t be afraid to mix and match


Download ppt "Introduction to Artificial Intelligence Revision Session."

Similar presentations


Ads by Google