Commit f4192fe

author

YuHang

committed

GTP use NN direct prediction

1 parent cb05575 commit f4192feCopy full SHA for f4192fe

File tree

2 files changed

+30

-18

lines changed

model
- APV_MCTS.py
- APV_MCTS_C.pyx

2 files changed

+30

-18

lines changed

`‎model/APV_MCTS.py‎`

Lines changed: 15 additions & 9 deletions

Original file line number	Diff line number	Diff line change
`@@ -183,7 +183,16 @@ def shift_node(self,move,pos_to_shift=None):`
`183`	`183`
`184`	`184`	`def suggest_move(self, position):`
`185`	`185`
`186`		`- move_prob = self.suggest_move_prob(position)`
	`186`	`+ """Use MCTS guided by NN"""`
	`187`	`+ #move_prob = self.suggest_move_prob(position)`
	`188`	`+`
	`189`	`+ """Use direct NN predition (pretty weak)"""`
	`190`	`+ move_probs,value = self.api.run_many(bulk_extract_features([position]))`
	`191`	`+ move_prob = move_probs[0]`
	`192`	`+ idx = np.argmax(move_prob)`
	`193`	`+ greedy_move = divmod(idx,go.N)`
	`194`	`+ prob = move_prob[idx]`
	`195`	`+ logger.debug(f'Greedy move is: {greedy_move} with prob {prob:.3f}')`
`187`	`196`
`188`	`197`	`on_board_move_prob = np.reshape(move_prob[:-1],(go.N,go.N))`
`189`	`198`	`if position.n < 30:`
`@@ -192,8 +201,11 @@ def suggest_move(self, position):`
`192`	`201`	`move = select_most_likely(position, on_board_move_prob)`
`193`	`202`
`194`	`203`	`player = 'B' if position.to_play==1 else 'W'`
`195`		`- win_rate = self.children[move].Q/2+0.5`
`196`		`- logger.info(f'Win rate for player {player} is {win_rate:2f}')`
	`204`	`+ """Use MCTS guided by NN average win ratio"""`
	`205`	`+ #win_rate = self.children[move].Q/2+0.5`
	`206`	`+ """Use direct NN value prediction (almost always 50/50)"""`
	`207`	`+ win_rate = value[0,0]/2+0.5`
	`208`	`+ logger.info(f'Win rate for player {player} is {win_rate:.4f}')`
`197`	`209`
`198`	`210`	`return move`
`199`	`211`
`@@ -206,12 +218,6 @@ def suggest_move_prob(self, position):`
`206`	`218`	`logger.debug(f'Expadning Root Node...')`
`207`	`219`
`208`	`220`	`move_probs,_ = self.api.run_many(bulk_extract_features([position]))`
`209`		`- '''`
`210`		`- move_prob = move_probs[0]`
`211`		`- greedy_move = divmod(np.argmax(move_prob),go.N)`
`212`		`- prob = move_prob[np.argmax(move_prob)]`
`213`		`- logger.debug(f'Greedy move is: {greedy_move} with prob {prob} at game step {position.n}')`
`214`		`- '''`
`215`	`221`	`self.position = position`
`216`	`222`	`self.expand(move_probs[0])`
`217`	`223`

`‎model/APV_MCTS_C.pyx‎`

Lines changed: 15 additions & 9 deletions

Original file line number	Diff line number	Diff line change
`@@ -183,7 +183,16 @@ class MCTSPlayerMixin(object):`
`183`	`183`
`184`	`184`	`def suggest_move(self, position):`
`185`	`185`
`186`		`- move_prob = self.suggest_move_prob(position)`
	`186`	`+ """Use MCTS guided by NN"""`
	`187`	`+ #move_prob = self.suggest_move_prob(position)`
	`188`	`+`
	`189`	`+ """Use direct NN predition (pretty weak)"""`
	`190`	`+ move_probs,value = self.api.run_many(bulk_extract_features([position]))`
	`191`	`+ move_prob = move_probs[0]`
	`192`	`+ idx = np.argmax(move_prob)`
	`193`	`+ greedy_move = divmod(idx,go.N)`
	`194`	`+ prob = move_prob[idx]`
	`195`	`+ logger.debug(f'Greedy move is: {greedy_move} with prob {prob:.3f}')`
`187`	`196`
`188`	`197`	`on_board_move_prob = np.reshape(move_prob[:-1],(go.N,go.N))`
`189`	`198`	`if position.n < 30:`
`@@ -192,8 +201,11 @@ class MCTSPlayerMixin(object):`
`192`	`201`	`move = select_most_likely(position, on_board_move_prob)`
`193`	`202`
`194`	`203`	`player = 'B' if position.to_play==1 else 'W'`
`195`		`- win_rate = self.children[move].Q/2+0.5`
`196`		`- logger.info(f'Win rate for player {player} is {win_rate:2f}')`
	`204`	`+ """Use MCTS guided by NN average win ratio"""`
	`205`	`+ #win_rate = self.children[move].Q/2+0.5`
	`206`	`+ """Use direct NN value prediction (almost always 50/50)"""`
	`207`	`+ win_rate = value[0,0]/2+0.5`
	`208`	`+ logger.info(f'Win rate for player {player} is {win_rate:.4f}')`
`197`	`209`
`198`	`210`	`return move`
`199`	`211`
`@@ -206,12 +218,6 @@ class MCTSPlayerMixin(object):`
`206`	`218`	`logger.debug(f'Expadning Root Node...')`
`207`	`219`
`208`	`220`	`move_probs,_ = self.api.run_many(bulk_extract_features([position]))`
`209`		`- '''`
`210`		`- move_prob = move_probs[0]`
`211`		`- greedy_move = divmod(np.argmax(move_prob),go.N)`
`212`		`- prob = move_prob[np.argmax(move_prob)]`
`213`		`- logger.debug(f'Greedy move is: {greedy_move} with prob {prob} at game step {position.n}')`
`214`		`- '''`
`215`	`221`	`self.position = position`
`216`	`222`	`self.expand(move_probs[0])`
`217`	`223`

0 commit comments

Comments

(0)

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Commit f4192fe

File tree

2 files changed

2 files changed

`‎model/APV_MCTS.py‎`

`‎model/APV_MCTS_C.pyx‎`

0 commit comments