artificial intelligence summary · summary: search & pathplanning usedin environments thatare...

24
ARTIFICIAL INTELLIGENCE Lecturer: Silja Renooij Summary Utrecht University The Netherlands These slides are part of the INFOB2KI Course Notes available from www.cs.uu.nl/docs/vakken/b2ki/schema.html INFOB2KI 2019-2020

Upload: others

Post on 12-Oct-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

ARTIFICIAL INTELLIGENCE

Lecturer: Silja Renooij

Summary

Utrecht University The Netherlands

These slides are part of the INFOB2KI Course Notes available fromwww.cs.uu.nl/docs/vakken/b2ki/schema.html

INFOB2KI 2019-2020

Page 2: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Subject overview Introduction (chapters 1,2) Pathfinding & Search (a.o. chapter 4) Learning (a.o. chapter 7) Deterministic planning & decision making (a.o. chapter 5, 6, 8, 11) Reasoning and planning under uncertainty (a.o. chapter 5) Movement (chapter 3)

2

Page 3: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

What is Intelligence? What is (game) AI?

Intelligence: A very general mental capability […] for comprehending our surroundings—"catching on," "making sense" of things, or "figuring out" what to do

Artificial Intelligence: four categories

Game AI: is about the illusion of human behaviour(Smart ‐‐ to a certain extent, unpredictable but rational,       emotional influences, body language to communicate  emotions, integrated in the environment)

Thinking rationally Thinking humanly

Acting rationally Acting humanly

3

Page 4: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Summary: Search & Pathplanning Used in environments that are static, deterministic, observable, and completely known(accessible)

Goal‐based Atomic states and actions (no domain structure) Uninformed search: BFS, DFS, UCS, IDS… Informed search: Greedy, A*,… Local search: hill‐climbing, simulated annealing, GAs,…

Adversarial search: minimax (alpha‐beta)

What aboutnatural paths?

4

Page 5: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Natural paths: Short…(?)Pre‐processing step:  automatic construction of waypoint graph, using sampling

Automatic construction of roadmaps for path planning in games, Nieuwenhuisen, Kamphuis, Mooijekind & Overmars, 2004.  5

Page 6: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Natural paths: enough clearance (?)Step 1: move waypoint c to ‘center’ cv

between obstacles

Idea: find closest point on obstacle move in opposite direction until

same distance to other obstacle

6

Page 7: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Natural paths: enough clearance (?)Step 2: move edges to ‘center’ between obstacles

Idea: Split edge with insufficient clearance  move middelpoint to center (green line) 

7

Page 8: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Natural paths: continuous (?)

Add a circular blend to every pair of incoming edges for every waypoint

Curvature depending on the amount of clearance

8

Page 9: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Natural paths result

9

Page 10: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Path planning:some problems remain…

Characters will still all follow the same path

No dynamic evasion No natural behavior Characters take turns that have no corresponding animation

In‐game changes to the characters have no effect on the path they follow Mood has no effect (although claimed by some games, there is no real evidence of it’s use) …

Openproblems

10

Page 11: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Summary: Learning Learning is essential for intelligence Learning is difficult when the world is dynamic, large and uncertain

Reinforcement learning does not necessarily need a model of the world

Supervised techniques learn from ‘examples’– Decision trees, Naïve Bayes classifiers– Neural Networks; also used as function approximators

Evolutionary algorithms used to let strategies/solutions compete  ~ search

11

Page 12: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Summary: Uncertainty

How to reason with ‘uncertain’ information– Fuzzy logic, Bayesian networks

What is the best strategy if– The outcome of actions is uncertain– The world state is uncertain

MDP, POMDP used to determine optimal actions (planning) under uncertainty

12

Page 13: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Summary Planning

Given initial state(s), goal(s) and set of actions: find a plan (a sequence of actions ) that is guaranteed to achieve goal(s). 

Inefficient as search: complete state descriptions, too many actions,  unused problem structure, multiple start and/or goal states,  does plan exist? 

Different planning formalisms for problems with different characteristics

13

Page 14: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Planning: problem characteristics Discrete (+ finite?) or continuous values Fully or partially observable One or more initial states

Deterministic or stochastic With or without duration Concurrent or sequential

Static or dynamic

Reach goal state or maximize reward

Single agent or multi‐agent– Cooperative or selfish– Individual or centralized planning

States:

Actions:

Env.:

Objective:

Planners:

14

Page 15: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Classical: STRIPS (GOAP), NOAH, HTN (SHOP)

Discrete (+ finite?) or continuous values Fully or partially observable 1 or more initial states

Deterministic or stochastic With or without duration Concurrent or sequential

Static or dynamic

Reach goal state or maximize reward

Single agent or multi‐agent– Cooperative or selfish– Individual or centralized planning

Dknown1

DWS

S

G

S

Except GOAP!

States:

Actions:

Env.:

Objective:

Planners:

15

Page 16: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Reactive: DT, FSM, BT, rule-based, …

Discrete (+ finite?) or continuous values Fully or partially observable 1 or more initial states

Deterministic or stochastic With or without duration Concurrent or sequential

Static or dynamic

Reach goal state or maximize reward

Single agent or multi‐agent– Cooperative or selfish– Individual or centralized planning

D

S

D

S

States:

Actions:

Env.:

Objective:

Planners:

16

Page 17: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Stochastic: MDP, POMDP

Discrete (+ finite?) or continuous values Fully or partially observable 1 or more initial states

Deterministic or stochastic With or without duration Concurrent or sequential

Static or dynamic

Reach goal state or maximize reward

Single agent or multi‐agent– Cooperative or selfish– Individual or centralized planning

Often D

≥ 1

SW

D

R

S

MDP: F, POMDP: PStates:

Actions:

Env.:

Objective:

Planners:

17

Page 18: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Summary Moving

Individual steering behaviors– Intelligent animation?– What if an animation fails halfway?

Group movement and teamwork– Crowd simulation:

• Move together but separate• Who stops first? Where do you stop? How do you pass obstacles?

18

Page 19: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

The MIRAnim Engine (for mixed reality)Cassell et al. identify two types of communicative body motions:

J. Cassell, T. Bickmore, M. Billinghurst, L. Campbell, K. Chang, Vilhjalmsson, H., and H. Yan. Embodiment in conversational interfaces: Rea. In Proceedings of the CHI’99 Conference, pages 520–527, 1999.

Intelligent Animation

Performance Animation

Interactive Virtual Humans

Gestures Posture shifts

MIRAnim

BLENDING control

idle

ness

19

Page 20: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Intelligent animation:

20

GRETA: Embodied Conversational AgentGreta can talk and simultaneously show facial expressions, gestures, gaze, and head movements. 

Page 21: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Intelligent Animation: problemsTo make agent realistic/believable: Blend of facial animation and body animation Blend of moving and handling objects Adjustment of moving/posture to environment Animation shouldn’t get “stuck” Not too realistic?

21

Page 22: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Not covered in this course

Constraint satisfaction Utility theory Natural Language Processing Vision/perception (recognizing objects) Knowledge representation (ontologies) Ethics …

22

Page 23: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Conclusions

Much is achieved:– Watson, Deep Blue– Alpha Go (Zero)– Robocup soccer– Autonomous cars– …

With every step taken new challenges become visible! (And not just in AI)

23

Page 24: ARTIFICIAL INTELLIGENCE Summary · Summary: Search & Pathplanning Usedin environments thatare static, deterministic, observable, andcompletelyknown (accessible) Goal‐based Atomic

Interested in more? Search the internet  Course page Other courses ICS dept:

– Applied Games (Ba IKU)– Kennissystemen (Ba IKU)– Intelligente Systemen (Ba ICA)– Computationele Intelligentie (Ba ICA)– Probabilistic Reasoning (Ma COSC + AI)– Evolutionary Computing (Ma COSC + AI)– AI for game technology (Ma GMT)– …

AI master (or COSC/GMT with electives)

Thank you!(please use Caracal forfeedback)

24