Presentation on theme: "11 Human Competitive Results of Evolutionary Computation Presenter: Mati Bot Course: Advance Seminar in Algorithms (Prof. Yefim Dinitz)"— Presentation transcript:
11 Human Competitive Results of Evolutionary Computation Presenter: Mati Bot Course: Advance Seminar in Algorithms (Prof. Yefim Dinitz)
22 Human-Competitiveness Definition Evolving Hyper-Heuristics using Genetic Programming Rush-Hour (bronze Humies prize in 2009 by Ami Hauptman) Freecell(gold Humies prize in 2011 by Achiya Elyasaf) Other Examples Human Competitive Results of Evolutionary Computation Outline
33 What is Human competitive? John Koza defined “Human-Competitiveness” in his book: Genetic Algorithms IV (2003). There are 8 criteria by which a result can be considered Human-Competitive.(will be explained in next slide) Our mission: Creation of Human-Competitive innovative solutions by means of Evolution.
44 The 8 Criteria of Koza for Human Competitivenes. (A) result is a Patent from the past, improvement of a patent. A New patent.
55 (B) result is equal to or better than another result that was published in a journal. The 8 Criteria of Koza for Human Competitivenes.
66 (C) result is equal to or better than a result in a known DB of results. The 8 Criteria of Koza for Human Competitivenes.
77 (D) publishable in its own right as a new scientific result. independent of the fact that the result was mechanically created. The 8 Criteria of Koza for Human Competitivenes.
88 (E) The result is equal to or better than the best human-created solution. The 8 Criteria of Koza for Human Competitivenes.
99 (F) equal to or better than an human achievement in its field at the time it was first discovered. The 8 Criteria of Koza for Human Competitivenes.
10 (G) The result solves a problem of indisputable difficulty in its field. The 8 Criteria of Koza for Human Competitivenes.
11 (H) The result holds its own or wins a regulated competition involving human contestants (in the form of either live human players or human-written computer programs). The 8 Criteria of Koza for Human Competitivenes.
12 Humies Competition “Humies annual Competition” gives $$$ for the best HC results. (in GECCO conference) Awarding a gold, silver and bronze prizes to the best entries. (money $$$) BGU won 1 gold, 1 silver and 6 bronze prizes since I counted more than 75 Human-Competitive results on the Humies competition site.
13 Evolving Hyper-Heuristics using Genetic Programming Ami Hauptman and Achiya Elyasaf
15 Representing Games as State-Graphs Every puzzle/game can be represented as a state graph: In puzzles, board games etc., every piece move can be counted as an edge/transition between states In computer war games etc. – the place of the player / the enemy, all the parameters (health, shield…) define a state
16 Rush-Hour as a state-graph Move blue Move purple
17 Searching Games State-Graphs Uninformed/naïve Search BFS – Breadth First Search Optimal solution Exponential space in the search depth DFS – Depth First Search(without node coloring). We might “never” track down the right path. Usually games contain cycles Linear Space Iterative Deepening: Combination of BFS & DFS Iterative Deepening Each iteration DFS with a depth limit is performed. Limit grows from one iteration to another Worst case - traverse the entire graph
18 Iterative Deepening
19 Searching Games State-Graphs Uninformed Search Most of the game domains are PSPACE- Complete! Worst case - traverse the entire graph We need an informed-search! (or an intelligent approach to traversing the graph)
20 Searching Games State-Graphs Heuristics Heuristic function h:states -> Real. For every state s, h(s) is an estimation of the minimal distance/cost from s to a solution In case h is perfect: an informed search that tries states with the lowest h-value first – will simply stroll to a solution For hard problems, finding a good h is hard Bad heuristic means the search might never track down the solution We need a good heuristic function to guide the informed search
21 Searching Games State-Graphs Informed Search Best-First search: Like DFS but select nodes with higher heuristic value first Best-First search Not necessarily optimal
22 Best-First Search
23 Searching Games State-Graphs Informed Search A*: A* G(s)=cost from root till s H(s)=Heuristic estimation F(s)=G(s)+H(s) Holds closed and sorted open lists(the list of states needs to be checked out). Best (=lowest F(s)) node of all open nodes is selected.
25 Searching Games State-Graphs Informed Search (Cont.) IDA*: Iterative-Deepening with A* IDA* The expanded nodes are pushed to the DFS stack by descending heuristic values Let g(s) be the cost to reach state s from root: Only nodes with f(s)=g(s)+h(s)
27 For H 1, …,H n – heuristics building blocks. How should we choose the fittest heuristic? Minimum? Maximum? Linear combination? GA/GP may be used for: Building new heuristics from existing building blocks Finding weights for each heuristic (for applying linear combination) Finding conditions for applying each heuristic Evolving Heuristics
28 Evolving Heuristics: GA W 1 =0.3W 2 =0.01W 3 =0.2…W n =0. 1
29 Evolving Heuristics: GP If And ≤ ≤ H1 0.4 ≥ ≥ H H2 * * H1 0.1 * * H5 / / H1 0.1 Condition True False
33 Rush Hour GP-Rush [Hauptman et al, 2009] Bronze Humies award
34 Domain-Specific Heuristics Hand-Crafted Heuristics / Guides: Blocker estimation – lower bound (admissible) Blocker estimation Goal distance – Manhattan distance Goal distance Hybrid blockers distance – combine the above two Hybrid blockers distance Is Move To Secluded – did the car enter a secluded area? (last move blocks all other cars) Is Move To Secluded Is a Releasing Move – if the last move increased the number of free cars.
35 Blockers Estimation Lower bound for number of steps to goal By: Counting moves to free blocking cars Example: O is blocking RED Need at least: Move O Move C Move B Move A H = 4
36 Goal Distance 16 Deduce goal Use “Manhattan Distance” from goal as h measure
39 Results Average reduction of nodes required to solve test problems, with respect to the number of nodes scanned by iterative deepening: H1: the heuristic function BlockersLowerBound. H2: GoalDistance. H3: Hybrid. Hc is our hand-crafted policy. GP is the best evolved policy, selected according to performance on the training set. Heuristic: Problem IDH1H2H3HcGP 6x6100%72%94%102%70%40% 8x8100%69%75%70%50%10%
40 Results (cont’d) Time (in seconds) required to solve problems JAM01... JAM40: ID – iterative deepening, Hi – average of our three hand-crafted heuristics, Hc – our hand-crafted policy. GP – our best evolved policy. human players (average of top 5).
41 FreeCell FreeCell remained relatively obscure until Windows 95 There are 32,000 solvable problems (known as Microsoft 32K), except for game #11982, which has been proven to be unsolvable Evolving hyper heuristic-based solvers for Rush-Hour and FreeCell [Hauptman et al, SOCS 2010] GA-FreeCell: Evolving Solvers for the Game of FreeCell [Elyasaf et al, GECCO 2011]
42 FreeCell (cont’d) As opposed to Rush Hour, blind search failed miserably The best published solver to date solves 96% of Microsoft 32K Reasons: High branching factor Hard to generate a good heuristic
43 Learning Methods: Random Deals Which deals (( חלוקות קלפים should we use for training? First method tested - random deals This is what we did in Rush Hour Here it yielded poor results Very hard domain
44 Learning Methods: Gradual Difficulty Second method tested - gradual difficulty Sort the problems by difficulty Each generation tests solvers against 5 deals from the current difficulty level + 1 random deal easyhard
45 A few words on Co-evolution Population 1Population 2 Test for fitness Solution, Solvers. Examples: Freecell Solver Rush Hour Solver Chess player Problems, adversaries, Examples: Freecell Deals Rush Hour Boards Another Chess Player Examples?
46 Learning Methods: Hillis-Style Co-evolution Third method tested - Hillis-style co-evolution using “Hall-of-Fame”: A deal population is composed of 40 deals (=40 individuals) + 10 deals that represent a hall-of- fame Each hyper-heuristic is tested against 4 deal individuals and 2 hall-of-fame deals
47 Learning Methods: Rosin-style Co-evolution Fourth method tested - Rosin-style co-evolution: Each deal individual consists of 6 deals Mutation and crossover: Crossover: Mutation p1 p p
48 Results Learning Method Run Node Reduction Time Reduction Length Reduction Solved -HSD100% 96% Gradual Difficulty GA-123%31%1%71% GA-227%30%103%70% GP---- Policy28%36%6%36% Rosin-style coevolution GA 87%93%41%98% Policy 89%90%40%99%
49 Other Human Competitive results Antenna Design for the International Space Station Automatically finding patches using genetic programming Evolvable Malware And many more on Humies site.