The End
+
AI @ RMIT
COSC1127/1125
Artificial Intelligence
Semester 2, 2021
Prof.
Wominjeka!
Week 12
AI @ RMIT
Prof.
@
Acknowledgement
I acknowledge the Traditional Owners (Woiwurrung and Boonwurrung form the Kulin nations) of the land on which we have conducted the whole Artificial Intelligence 2021 course during this semester.
I, and invite you all to, pay our respects to their Elders, past and present, and the first people of other communities who may be here today. And to be aware of the challenges and injustices that unfortunately continue up to our times.
I recognise their continuing connection to land, waters and culture.
It’s a way of showing awareness of, and respect for, the Aboriginal Traditional Owners of the land where a meeting or event is held.
‹#›
Some news…
Bonus Project 3 results sent to all
THE Review done last week
Final Pacman Contest:
Agent system due this Week 12
Wiki + Video due next Week 13
Instructions on Wiki and video and Wiki template to be provided tomorrow Wednesday on forum.
Final THE: Week 15
CES Survey closes soon?
Are you a better CS after this course?
Have you learnt new things as a CS?
*
Project 3: RL
Project 3 feedback
Reinforcement Learning is now my most favorite thing
I’m not sure why this was a bonus assessment – it was very useful for learning how the RL concepts worked in an applied fashion
I do like that this project was a bonus. Enjoyed what I did without stressing about needing to get above 50%. Best balance of learning / reward / still not counting for too much or being overly important.
This was the most awesome challenge yet… to learn this kind of knowledge is exactly why I signed up for
Some interesting techniques which I would like to implement into my Pacman game project.
Quite interesting. Very useful. Got a better understanding of RL (it doesn’t feel so scary anymore). Enjoyed it very much.
Wiki + Video
Due end of Week 13 both: 10% + 10%
Tell us about your system:
How does your agent system work?
What you tried and didn’t work; why?
Performance evaluation and analysis; reflections on what you tried (and worked and did not worked)
Mere plain statements (“I tried X and did not work”) will not be considered towards mark.
Wiki as part of your repo.
Template to be provided tomorrow.
Video to be uploaded to youtube and linked from Wiki.
Everybody must present.
Week 15 Final THE
20% of course marks; individual assessment.
Thursday Nov 4th @ 9am-9pm (12hrs!)
2-3hrs long test; many targeted questions.
Link will be posted in EdStem.
Be ready: NO REPLACEMENT without SPC.
SPC as an Equivalent Assessment on paper and on campus with me later (~December).
Submit before 6pm when silence policy starts (don’t risk it!)
Two parts:
PDF document with question specs.
Online Google Form to enter answers. Make 100% sure you are logged to your RMIT account.
You can have your own pen & paper notes.
Similar to Week 9 THE but covering ALL weeks.
Content: Weeks 1 – 12 (included).
Theoretical & conceptual questions: ~tutorials.
No negative marking.
Development of Project Contest
Questions?
*
Later….
@
Click image!
‹#›
It provides the core knowledge of computer science.
Hence, serious programmers and software engineers should know about major AI techniques.
It can be profitable.
It is also fun.
It relates to other non-exact sciences.
It is different to most other subjects.
Why AI is an important course?
*
What have we seen?
Search: as a general problem technique.
Knowledge Representation: rational behavior requires knowledge! Beyond databases…
Automated Planning: what plan should I execute?
mixing search + KR
MDP: decision making under uncertainty
Reinforcement Learning: learn environment and how to act rationally.
Probabilities: basic tool for reasoning under uncertainty.
Bayesian Networks: knowledge representation for probabilistic reasoning.
*
So, what is “AI”?
“Artificial intelligence is that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment.” — Nils J. Nilsson
You say more?
Thus we try to construct systems whose computation achieves or approximates the desired notion of rationality:
Rationality: Precise mathematical notion of what it means to do the right thing in any particular circumstance.
How to achieve “intelligent” behavior through computational means.
Four AI Definitions by R&N
Like Humans Not necessary like humans
Think Systems that think like humans
Systems that think rationally
Act Systems that act like humans Systems that act rationally
Cognitive Science
OUR FOCUS
Artificial Intelligence
Course objectives:
`Basic concepts and techniques of AI’
Search
Knowledge representation
Planning
Intelligent Agents
Reasoning under uncertainty
Machine Learning
*
Search
*
Search is an exploration of possibilities.
Particularly useful when a ‘direct’ method/algorithm is not known: extremely general!
One of the first topics studied in AI:
Newell and Simon (1961) General Problem Solver.
Central component to many AI systems:
Automated reasoning, theorem proving, robot navigation, VLSI layout, scheduling, game playing,…
What is search?
Search as graph reachability
Initial state
Goal states
*
Search Problems
A search problem is defined by:
Possible states
Initial state
Actions
Transition model
Goal test
Path cost
Initial State
Goal State
TRANSFORM INTO
How to do this?
*
Game Tree Search
MAX NODE
MIN NODE
TERMINALS
MIN ACTIONS
MAX ACTIONS
Representing knowledge & reasoning
Logic, Proofs, Databases & AI KRR
‹#›
Logic, Proofs, Databases & AI KRR
QUERY:
Is ɸ true?
ɸ1, ɸn, …, ɸn,
‹#›
*
Semantics
Entailment: KB |= α
α is true in all the models in which KB is true
models
KB
α
models
KB
α
models
KB
α
KB |= α
KB |≠ α
KB |≠ α
Entailment & Satisfiability
KB ╞ A
iff
interpetations that make KB true, make A true
(all models of KB are models of A )
iff
there is no model of KB ∧ ¬A
iff
KB ∧ ¬A is unsatisfiable
RESOLUTION
*
Resolution: An Example
Want to prove (P ∨ Q) ∧ (¬P ∨ Q) ╞ (R ∨ Q)
Prove (P ∨ Q) ∧ (¬P ∨ Q) ∧ ¬(R ∨ Q) is unsatisfiable.
Prove (P ∨ Q) ∧ (¬P ∨ Q) ∧ (¬R) ∧ (¬Q) is unsat.
Apply resolution:
(P ∨ Q)
(¬R)
(¬Q)
(¬P ∨ Q)
(Q)
()
Contradiction!!
*
First order logic: Basic Elements
Objects: basic individuals such as people, numbers, colours, games, chess pieces, ….
wqueen, door23, mary, table, block7, etc.
Functions: map objects to objects
motherOf(mary), age(jim), plus(2,3), president(usa), …
Variables: placeholders for objects
x, y, z, …
Terms: complex denotation of individuals build from functions, objects, and variables:
motherOf(fatherOf(mary)), plus(mult(2,5),plus(4,8))
motherOf(x), plus(x,minus(y,2)), age(fatherOf(x))
*
Reasoning under Uncertainty
“When it is not in our power to determine what is true, we ought to follow what is most probable.” ― René Descartes
Theoretical reasoning:
use of reason to decide what to believe.
Practical reasoning:
use of reason to decide how to act.
*
The Axioms of Probability
For any random variable A: 0 ≤ P(A) ≤ 1
P(True) = 1 (you can see True = A v -A)
P(False) = 0 (you can see False = A Λ -A)
P(A v B) = ???
P(A v B) = P(A) + P(B) – P(A Λ B)
These axioms limit the class of functions that can be considered as probability functions
Bayes Rule
P(A|B)P(B) = P(AΛB) = P(BΛA) = P(B|A)P(A)
Bayes Rule:
P(B|A)= [P(A|B)P(B)]/P(A)
Know this!!
Motivation: when calculating P(B|A) we often know distribution P(A|B). For example:
A represents symptoms evidences
B represents illness or condition
Bayesian Networks
Burglary Network
Query: P(B = true | A = true, M = false, J = false)?
… or, what is the same: P(b | a, m, ¬j)?
Autonomous Behavior in AI:
The Control Problem
The key problem is to select the action to do next.
Three approaches to this problem:
Programming-based: Specify control by hand.
Learning-based: Learn control from experience.
Model-based: Specify problem by hand, derive control automatically.
Approaches orthogonal though; and successes and limitations in each . . .
the control problem
*
Automated Planning…
International Conference on Automated Planning and Scheduling
How to develop systems or ’agents’
that can make decisions on their own?
*
Solution 3: Model-based Approach
Specify model for problem:
actions, initial situation, goals, and sensors
Let a solver compute controller automatically!
Advantage: flexible, clear, and domain-independent.
Disadvantage: need a model; computationally intractable.
Model-based approach to intelligent behavior: AI Planning.
SOLVER
Controller
actions
World
observations
Action
Sensors Goals
*
Classical Planning
PLANNER
Sequential Plan
State Model
Language
Algorithm
Graphplan-based Heuristics
Build reachability graph P0, A0, P1, A1, . . .
P0 = {p ∈ Init}
Ai = {a ∈ O | Prec(a) ⊆ Pi}
Pi+1 = Pi ∪ {p ∈ Add(a) | a ∈ Ai}
Graph implicitly represents max heuristic:
hmax(s) = min i such that G ⊆ Pi
P0
A0
P1
A1
Pk
*
MDP + RL: Acting under uncertainty
Say something about the course here:
2580 7725 @ menti.com
Time for…
@
Click image!
What have we seen in this course?
Search: as a general problem technique.
Knowledge Representation: rational behavior requires knowledge! Beyond databases…
Probabilities: basic tool for reasoning under uncertainty.
Automated Planning: what plan should I execute?
mixing search + KR
MDP: decision making under uncertainty
Reinforcement Learning: learn environment and how to act rationally.
Bayesian Networks: knowledge representation for probabilistic reasoning.
*
Side effects of taking AI…
Learn Python or get better at it.
Learn better development practices (including git!)
Learn about LaTex.
Improve teamwork skills.
Make friends and future colleagues!
The Questions
What would you tell a fellow student who is interested in the course?
Do you think you will be a better Computer Scientist after taking AI?
Two last videos to discuss…
Prof.
On the Turing Test
Prof.
Ethical Crisis in CS?