ARTIFICIAL INTELLIGENCE Summary Lecturer: Silja Renooij These - PowerPoint PPT Presentation

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

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

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 Thinking rationally Thinking humanly Acting rationally Acting humanly  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) 3

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… What about  Informed search: Greedy, A*,… natural paths?  Local search: hill‐climbing, simulated annealing, GAs,…  Adversarial search: minimax (alpha‐beta) 4

Natural paths: Short…(?) Pre‐processing step:  automatic construction of waypoint graph, using sampling Automatic construction of roadmaps for path planning in games , 5 Nieuwenhuisen, Kamphuis, Mooijekind & Overmars, 2004.

Natural paths: enough clearance (?) Step 1: move waypoint c to ‘center’ c v between obstacles Idea:  find closest point on obstacle  move in opposite direction until same distance to other obstacle 6

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

Natural paths: continuous (?)  Add a circular blend to every pair of incoming edges for every waypoint  Curvature depending on the amount of clearance 8

Natural paths result 9

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 Open no effect on the path they follow problems  Mood has no effect (although claimed by some games, there is no real evidence of it’s use)  … 10

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

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

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

Planning: problem characteristics States:  Discrete (+ finite?) or continuous values  Fully or partially observable  One or more initial states Actions:  Deterministic or stochastic  With or without duration  Concurrent or sequential Env.:  Static or dynamic Objective:  Reach goal state or maximize reward Planners:  Single agent or multi‐agent – Cooperative or selfish – Individual or centralized planning 14

Classical: STRIPS (GOAP), NOAH, HTN (SHOP)  Discrete (+ finite?) or continuous values States: D  Fully or partially observable known  1 or more initial states 1 Actions:  Deterministic or stochastic D  With or without duration W  Concurrent or sequential S Env.:  Static or dynamic S Objective:  Reach goal state or maximize reward G  Single agent or multi‐agent Planners: S – Cooperative or selfish Except GOAP! – Individual or centralized planning 15

Reactive: DT, FSM, BT, rule-based, …  Discrete (+ finite?) or continuous values States:  Fully or partially observable  1 or more initial states Actions:  Deterministic or stochastic D  With or without duration  Concurrent or sequential S Env.:  Static or dynamic D Objective:  Reach goal state or maximize reward  Single agent or multi‐agent Planners: S – Cooperative or selfish – Individual or centralized planning 16

Stochastic: MDP, POMDP  Discrete (+ finite?) or continuous values States: Often D  Fully or partially observable MDP: F, PO MDP: P ≥ 1  1 or more initial states Actions:  Deterministic or stochastic S  With or without duration W  Concurrent or sequential Env.:  Static or dynamic D Objective:  Reach goal state or maximize reward R  Single agent or multi‐agent Planners: S – Cooperative or selfish – Individual or centralized planning 17

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

Intelligent Animation The MIRAnim Engine (for mixed reality) Cassell et al. identify two types of communicative body motions: Gestures Posture shifts idleness MIRAnim Performance Animation Interactive Virtual Humans BLENDING control 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. 19

Intelligent animation: GRETA: Embodied Conversational Agent Greta can talk and simultaneously show facial expressions, gestures, gaze, and head movements. 20

Intelligent Animation: problems To 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

Not covered in this course  Constraint satisfaction  Utility theory  Natural Language Processing  Vision/perception (recognizing objects)  Knowledge representation (ontologies)  Ethics  … 22

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

Interested in more?  Search the internet  Thank you!  Course page (please use Caracal for feedback)  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) 24