-
Notifications
You must be signed in to change notification settings - Fork 1
/
core.clj
277 lines (220 loc) · 8.83 KB
/
core.clj
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
(ns jc-clj-mcts.core)
;;MCTS in clojure
;;http://randomcomputation.blogspot.com/2013/01/monte-carlo-tree-search-in-clojure.html
;;;================================TTT
(def wins [[0 1 2] [3 4 5] [6 7 8] ;cols
[0 3 6] [1 4 7] [2 5 8] ;rows
[0 4 8] [2 4 6]]) ;diags
(defn opp-player [p] (if (= p :X) :O :X))
(defn blank? [x] (number? x))
(defn win-check [state player]
(loop [wins wins]
(if (empty? wins) false
(let [[a b c] (first wins)]
(if (and (= player (get state a))
(= player (get state b))
(= player (get state c)))
true
(recur (rest wins)))))))
(defn ttt-terminal? [state]
(if (win-check state :X) {:draws 0 :x-win 1 :o-win 0}
(if (win-check state :O) {:draws 0 :x-win 0 :o-win 1}
(if (not-any? blank? state) {:draws 1 :x-win 0 :o-win 0} false))))
(defn ttt-gen-children [state to-move]
(for [[i v] (zipmap (range) state)
:when (blank? v)]
(assoc state i to-move)))
(defn ttt-playout
"Returns value of playout simulation for tic tac toe"
[state to-move]
(loop [state state to-move to-move]
(if-let [result (ttt-terminal? state)] result
(recur (rand-nth (ttt-gen-children state to-move))
(opp-player to-move)))))
;;;================================UCT
(def init-record
"The data associated with a board-state in the mem"
{:visits 0 :draws 0
:x-win 0 :o-win 0
:chldn [] :to-move :X})
(def init-mem
"Creates initial mem for TTT"
(let [init-state [ 0 0 0
0 0 0
0 0 0]]
{init-state
(assoc init-record :chldn (ttt-gen-children init-state :X))}))
(defn uct-value
"Value of a state based on gathered statistics. Currently not
actually 'uct' value (see paper)."
[{:keys [visits x-win o-win draws to-move]}]
(case (opp-player to-move)
:X (/ (+ x-win (/ draws 2))
visits)
:O (/ (+ o-win (/ draws 2))
visits)
:default 0))
(defn uct-sample
"The random sampling function for a board state."
[state mem func times]
(loop [result {:draws 0 :x-win 0 :o-win 0} times times]
(if (< times 1) result
(recur (reduce
(fn [m [k v]]
(update-in m [k] + v))
result
(func state (get-in mem [state :to-move])))
(dec times)))))
(defn uct-select
"Selects highest rated child of state"
[mem state]
(let [chldn (get-in mem [state :chldn])
explrd (remove
#(zero? (get-in mem [% :visits] 0))
chldn)]
(if (empty? explrd)
(rand-nth chldn)
(apply max-key #(uct-value (get mem %)) explrd))))
(defn uct-unexplored [mem state]
"Unexplored children of state"
(for [c (get-in mem [state :chldn]
(ttt-gen-children state (get-in mem [state :to-move])))
:when (zero? (get-in mem [c :visits] 0))] c))
(defn uct-backprop
"Backpropagates child value to the parent"
[mem path {:keys [x-win o-win draws] :as stats}]
(if (empty? path) mem
(recur
(-> mem
(update-in [(first path) :x-win] + x-win)
(update-in [(first path) :o-win] + o-win)
(update-in [(first path) :draws] + draws)
(update-in [(first path) :visits] inc))
(rest path)
stats)))
(defn- add-child
"Helper to creates child-record for the mem."
[mem parent-state child-state]
(let [to-move (get-in mem [parent-state :to-move])
child-record (get mem
child-state
(assoc init-record
:chldn (ttt-gen-children
child-state
(opp-player to-move))
:to-move (opp-player to-move)))]
(assoc mem child-state child-record)))
(defn uct-grow
"Estimates a child's value and adds it to the tree."
[mem path]
(let [leaf (first path)
chld (rand-nth (uct-unexplored mem leaf))
valu (uct-sample chld mem ttt-playout 1)]
(-> mem
(add-child leaf chld)
(uct-backprop (cons chld path) valu))))
(defn learn-iteration [mem state]
"The core algorithm; a single analysis of a state. Searches the tree
for an unexplored child, estimates the child's value, adds
it to the tree, and backpropagates the value up the path."
(loop [mem mem, state state, path (list state)]
(if-let [result (ttt-terminal? state)]
(uct-backprop mem path result)
(if (not-empty (uct-unexplored mem state))
(uct-grow mem path)
(let [ch (uct-select mem state)]
(recur mem ch (cons ch path)))))))
(defn learn-state [mem state budget]
"Analyzes a board state using the UCT algorithm. Iterates
learn-iteration until budget is exhausted."
(loop [mem mem budget budget]
(if (< budget 1) mem
(recur (learn-iteration mem state) (dec budget)))))
;;;================================RUN
(defn print-board [board]
"Pretty print a board state"
(println
(apply format "%s %s %s \n%s %s %s \n%s %s %s \n"
(map #(case % :X "X" :O "O" 0 "_") board))))
(defn play-game
"Retains memory built from analyses of past moves"
[[mem _]]
(let [uctp (rand-nth [:X :O])]
(loop [mem mem
board-state [0 0 0 0 0 0 0 0 0]
to-move :X]
(if-let [{:keys [draws x-win o-win]} (ttt-terminal? board-state)]
[mem
(hash-map
:uct (if (= uctp :X) x-win o-win)
:rnd (if (= uctp :X) o-win x-win)
:draws draws)]
(if (= uctp to-move)
(let [mem (learn-state mem board-state 30)
move (uct-select mem board-state)]
(recur mem move (opp-player to-move)))
(let [move (rand-nth (get-in mem [board-state :chldn]))
mem (if (contains? mem move) mem
(assoc mem move
(assoc init-record
:chldn (ttt-gen-children
move
(opp-player to-move))
:to-move (opp-player to-move))))]
(recur mem move (opp-player to-move))))))))
(defn play-game-no-mem
"Does not retain memory over moves to allow for
effectiveness assessment based on computational budget"
[budget]
(let [uctp (rand-nth [:X :O])]
(loop [board-state [0 0 0 0 0 0 0 0 0]
to-move :X]
; (print-board board-state)
(if-let [{:keys [draws x-win o-win]} (ttt-terminal? board-state)]
(hash-map
:uct (if (= uctp :X) x-win o-win)
:rnd (if (= uctp :X) o-win x-win)
:draws draws)
(if (= uctp to-move)
(let [mem (learn-state (hash-map
board-state
(assoc init-record
:to-move to-move
:chldn (ttt-gen-children
board-state
to-move)))
board-state
budget)
move (uct-select mem board-state)]
(recur move (opp-player to-move)))
(let [move (rand-nth (ttt-gen-children board-state to-move))]
(recur move (opp-player to-move))))))))
(defn- update-stats
"Helper function"
[curr new]
(reduce
(fn [m [k v]] (update-in m [k] + v))
curr new))
(defn uct-v-rand
"Plays n games of uct vs rand retaining the analysis memory
across games"
[n]
(loop [mem init-mem
games 0
stats {:uct 0 :rnd 0 :draws 0}]
(if (> games n) [mem stats]
(let [[mem result] (play-game mem)]
(recur mem
(inc games)
(update-stats stats result))))))
;;;This script generates the results table in the blog post
;(let [data (for [b [0 1 2 3 4 5 10 100]] ;computational budgets
; (list b (take 50 (repeatedly #(play-game-no-mem b)))))
; stats (map
; (fn [[b d]]
; (let [avgs {:uct (float (/ (reduce + (map :uct d)) (count d)))
; :rnd (float (/ (reduce + (map :rnd d)) (count d)))
; :draws (float (/ (reduce + (map :draws d)) (count d)))}]
; (list b avgs)))
; data)]
; (pprint stats))