AI2GTP

CONTRIBUTORS

@@ -1,3 +1,7 @@
 Primary author and project maintainer (https://github.com/sanderland/katrain):
 Sander Land
 
+Thanks to:
+'Dontbtme' for feedback and testing of v1.0.
+
+

TODO

@@ -10,15 +10,14 @@
 [x] Self-play tournaments in separate script.
 [x] ai thoughts in sgf
 [x] more AI modes?
+[x] pol value override > 0.95 ?  policy threshold override p+pick?
+
 
 [/] README
 [] engine status
-[] policy threshold override p+pick?
+[] sgf review improvements  --  Likewise, in the 0.9 version, better alternatives to the played move were shown with squares, which was also pretty useful when using the sgf outside of Katrain. I mean, having the top move mentioned is all and good, but when you see multiple squares shown on the board as better alternatives to the move played in the game, it makes obvious how far from perfect that move actually was :D
-[] sgf review improvements
 [] Release notes
- --  Likewise, in the 0.9 version, better alternatives to the played move were shown with squares, which was also pretty useful when using the sgf outside of Katrain. I mean, having the top move mentioned is all and good, but when you see multiple squares shown on the board as better alternatives to the move played in the game, it makes obvious how far from perfect that move actually was :D
 
-[] pol value override > 0.9 ?
 [] clarify score change vs score
 [] pv with overlap?
 [] config player to sep. row/popups?

ai.py

@@ -8,7 +8,7 @@ import numpy as np
 
 from common import OUTPUT_INFO, var_to_grid, OUTPUT_DEBUG, OUTPUT_ERROR
 from engine import EngineDiedException
-from game import Move, Game, IllegalMoveException
+from game import Move, Game, IllegalMoveException, GameNode
 
 
 def weighted_selection_without_replacement(items: List[Tuple[float, float, int, int]], pick_n: int) -> List[Tuple[float, float, int, int]]:
@@ -25,7 +25,7 @@ def fmt_moves(moves: List[Tuple[float, Move]]):
     return ", ".join(f"{mv.gtp()} ({p:.2%})" for p, mv in moves)
 
 
-def ai_move(game: Game, ai_mode: str, ai_settings: Dict):
+def ai_move(game: Game, ai_mode: str, ai_settings: Dict) -> Tuple[Move, GameNode]:
     cn = game.current_node
     while not cn.analysis_ready:
         time.sleep(0.01)
@@ -69,11 +69,11 @@ def ai_move(game: Game, ai_mode: str, ai_settings: Dict):
             n_moves = int(ai_settings["pick_frac"] * len(legal_policy_moves) + ai_settings["pick_n"])
             if "influence" in ai_mode or "territory" in ai_mode:
                 if "influence" in ai_mode:
-                    weight = lambda x, y: ai_settings["influence_weight"] ** max(0, 3 - min(size[0] - 1 - x, x, y, size[1] - 1 - y))
+                    weight = lambda x, y: ai_settings["line_weight"] ** max(0, 3 - min(size[0] - 1 - x, x, y, size[1] - 1 - y))
                 else:
-                    weight = lambda x, y: ai_settings["influence_weight"] ** max(0, min(size[0] - 1 - x, x, y, size[1] - 1 - y) - 2)
+                    weight = lambda x, y: ai_settings["line_weight"] ** max(0, min(size[0] - 1 - x, x, y, size[1] - 1 - y) - 2)
                 weighted_coords = [(policy_grid[y][x] * weight(x, y), weight(x, y), x, y) for x in range(size[0]) for y in range(size[1]) if policy_grid[y][x] > 0]
-                ai_thoughts += f"Generated weights for {ai_mode} according to weight factor {ai_settings['influence_weight']} and distance from 4th line. "
+                ai_thoughts += f"Generated weights for {ai_mode} according to weight factor {ai_settings['line_weight']} and distance from 4th line. "
             elif "local" in ai_mode or "tenuki" in ai_mode:
                 var = ai_settings["local_stddev"] ** 2
                 if not cn.single_move or cn.single_move.coords is None:
@@ -100,8 +100,8 @@ def ai_move(game: Game, ai_mode: str, ai_settings: Dict):
                 aimove = new_top[0][1]
                 ai_thoughts += f"Top 5 among these were {fmt_moves(new_top)} and picked top {aimove.gtp()}. "
                 if new_top[0][0] < pass_policy:
-                    ai_thoughts += f"But found pass ({pass_policy:.2%} to be higher rated than {aimove.gtp()} ({new_top[0][0]:.2%}) so will pass instead."
+                    ai_thoughts += f"But found pass ({pass_policy:.2%} to be higher rated than {aimove.gtp()} ({new_top[0][0]:.2%}) so will play top policy move instead."
-                    aimove = Move(None, player=cn.next_player)
+                    aimove = top_policy_move
             else:
                 aimove = top_policy_move
                 ai_thoughts += f"Pick policy strategy {ai_mode} failed to find legal moves, so is playing top policy move {aimove.gtp()}."
@@ -113,12 +113,8 @@ def ai_move(game: Game, ai_mode: str, ai_settings: Dict):
             move
             for i, move in enumerate(candidate_ai_moves)
             if i == 0
-            or move["visits"] >= ai_settings["balance_min_visits"]
+            or move["visits"] >= ai_settings["min_visits"]
-            and (
+            and (move["pointsLost"] < ai_settings["random_loss"] or move["pointsLost"] < ai_settings["max_loss"] and sign * move["scoreLead"] > ai_settings["target_score"])
-                move["pointsLost"] < ai_settings["balance_random_loss"]
-                or move["pointsLost"] < ai_settings["balance_max_loss"]
-                and sign * move["scoreLead"] > ai_settings["balance_target_score"]
-            )
         ]
         aimove = Move.from_gtp(random.choice(sel_moves)["move"], player=cn.next_player)  # TODO: could be weighted towards worse
         ai_thoughts += f"Balance strategy selected moves {sel_moves} based on target score and max points lost, and randomly chose {aimove.gtp()}."
@@ -137,7 +133,6 @@ def ai_move(game: Game, ai_mode: str, ai_settings: Dict):
     try:
         played_node = game.play(aimove)
         played_node.ai_thoughts = ai_thoughts
+        return aimove, played_node
     except IllegalMoveException as e:
         game.katrain.log(f"AI Strategy {ai_mode} generated illegal move {aimove.gtp()}:  {e}", OUTPUT_ERROR)
-
-    return aimove, played_node

ai2gtp.py

@@ -35,22 +35,25 @@ ai_settings = {
     "noise_strength": 0.8,
     "pick_n": 10,
     "pick_frac": 0.2,
-    "local_stddev": 10,
+    "stddev": 10,
-    "influence_weight": 0.1,
+    "line_weight": 0.1,
     "pick_override": 0.95,
 }
 
 engine = KataGoEngine(logger, ENGINE_SETTINGS)
-ai_strategy = "P+Pick"
-ai_strategy = "P+Local"
-ai_settings["pick_frac"] = 0.0
-ai_settings["pick_n"] = 10
-ai_settings["local_stddev"] = 1.0
 
 ai_strategy = "P+Influence"
 ai_settings["pick_frac"] = 0.5
-ai_settings["influence_weight"] = 0.1
+ai_settings["line_weight"] = 0.1
 
+ai_strategy = "P+Local"
+ai_settings["pick_frac"] = 0.0
+ai_settings["pick_n"] = 25
+ai_settings["stddev"] = 1.5
+
+ai_strategy = "P+Pick"
+ai_settings["pick_frac"] = 0.33  # dropping below 7k
+ai_settings["pick_n"] = 5  # dropping below 7k at 5/0.33
 
 logger.log(f"STARTED ENGINE", OUTPUT_ERROR)
 
@@ -70,6 +73,7 @@ while not game.ended:
         logger.log(f"Setting komi {game.root.properties}", OUTPUT_ERROR)
     elif "genmove" in line:
         game.current_node.analyze(engine)
+        game.root.add_property(f"P{game.current_node.next_player}", f"KaTrain {ai_strategy}")
         move, node = ai_move(game, ai_strategy, ai_settings)
         logger.log(f"SENT TO GTP: = {move.gtp()}", OUTPUT_ERROR)
         print(f"= {move.gtp()}\n")
@@ -81,11 +85,10 @@ while not game.ended:
         moves = sorted(list(cn.analysis["moves"].values()), key=lambda d: d["order"])
         if moves:
             pv = " ".join(moves[0]["pv"])
-        # print(cn.analysis,cn.analysis.get('root'),file=sys.stderr)
         print(
             f"CHAT:Visits {cn.ai_thoughts} Winrate {cn.analysis['root']['winrate']:.2%} ScoreLead {cn.analysis['root']['scoreLead']:.1f} ScoreStdev 0.0 PV {move.gtp()} {pv}",
             file=sys.stderr,
-        )  #
+        )
         continue
     elif "play" in line:
         _, player, move = line.split(" ")

config.json

@@ -38,16 +38,14 @@
         "eval_show_ai": false
     },
     "ai": {
-        "balance_target_score": 2,
+        "Default": {},
-        "balance_random_loss": 1,
+        "Balance": {"target_score":2, "random_loss": 1, "max_loss": 5, "min_visits": 20},
-        "balance_max_loss": 5,
+        "P+Noise" : {"noise_strength": 0.8},
-        "balance_min_visits": 20,
+        "P+Pick": {"pick_override": 0.95, "pick_n":5, "pick_frac": 0.33},
-        "noise_strength": 0.8,
+        "P+Local": {"pick_override": 0.95, "stddev": 1.5, "pick_n":15, "pick_frac": 0.0},
-        "pick_override": 0.95,
+        "P+Tenuki": {"pick_override": 0.95, "stddev": 10, "pick_n":5, "pick_frac": 0.33},
-        "pick_n": 10,
+        "P+Influence": {"pick_override": 0.95, "pick_n":5, "pick_frac": 0.5,"line_weight": 0.1},
-        "pick_frac": 0.2,
+        "P+Territory": {"pick_override": 0.95, "pick_n":5, "pick_frac": 0.5,"line_weight": 0.1}
-        "local_stddev": 5,
-        "influence_weight": 0.1
     },
     "board_ui": {
         "starpoint_size": 0.1,

game.py

@@ -34,7 +34,7 @@ class Game:
         if move_tree:
             self.root = move_tree
             self.komi = self.root.komi
-            handicap = int(self.root.get_first("HA",0))
+            handicap = int(self.root.get_first("HA", 0))
             if handicap and not self.root.placements:
                 self.place_handicap_stones(handicap)
         else:

gui/controls.py

@@ -1,6 +1,9 @@
 from kivy.graphics.context_instructions import Color
 from kivy.graphics.vertex_instructions import Line, SmoothLine
 from kivy.uix.boxlayout import BoxLayout
+from kivy.uix.popup import Popup
+
+from gui.popups import ConfigAIPopup
 
 
 class Controls(BoxLayout):
@@ -31,7 +34,7 @@ class Controls(BoxLayout):
         return self.player_mode_groups[player].value
 
     def ai_mode(self, player):
-        return self.ai_mode_groups[player].text.lower()
+        return self.ai_mode_groups[player].text
 
     def on_size(self, *args):
         self.update_evaluation()
@@ -60,13 +63,19 @@ class Controls(BoxLayout):
                 self.score.text = current_node.format_score()
                 self.win_rate.text = current_node.format_win_rate()
                 if move and next_player_is_human_or_both_robots:  # don't immediately hide this when an ai moves comes in
-                    self.score_change.label = f"Score change"
                     points_lost = current_node.points_lost
-                    self.score_change.text = f"{move.player}{-current_node.points_lost:+.1f}" if points_lost else "..."
+                    self.score_change.label = f"Points lost for {move.player}" if points_lost and points_lost > 0 else f"Points gained for {move.player}"
+                    self.score_change.text = f"{abs(points_lost):.1f}" if points_lost else "..."
                 elif not current_player_is_ai_playing_human:
-                    self.score_change.label = f"Score change"
+                    self.score_change.label = f"Points lost"
                     self.score_change.text = ""
 
             self.graph.update_value(current_node)
 
         self.info.text = info
+
+    def configure_ais(self):
+        config_popup = Popup(title="Edit AI Settings", size_hint=(0.9, 0.9))
+        popup_contents = ConfigAIPopup(self.katrain, config_popup, {self.ai_mode("B"), self.ai_mode("W")})
+        config_popup.add_widget(popup_contents)
+        config_popup.open()

gui/popups.py

@@ -23,6 +23,19 @@ class QuickConfigGui(BoxLayout):
         if initial_values:
             self.set_properties(self, initial_values)
 
+    @staticmethod
+    def type_to_widget_class(value):
+        if isinstance(value, float):
+            return LabelledFloatInput
+        elif isinstance(value, bool):
+            return LabelledCheckBox
+        elif isinstance(value, int):
+            return LabelledIntInput
+        if isinstance(value, dict):
+            return LabelledObjectInputArea
+        else:
+            return LabelledTextInput
+
     def collect_properties(self, widget):
         if isinstance(widget, (LabelledTextInput, LabelledSpinner, LabelledCheckBox)):
             try:
@@ -69,19 +82,6 @@ class NewGamePopup(QuickConfigGui):
 
 
 class ConfigPopup(QuickConfigGui):
-    @staticmethod
-    def type_to_widget_class(value):
-        if isinstance(value, float):
-            return LabelledFloatInput
-        elif isinstance(value, bool):
-            return LabelledCheckBox
-        elif isinstance(value, int):
-            return LabelledIntInput
-        if isinstance(value, dict):
-            return LabelledObjectInputArea
-        else:
-            return LabelledTextInput
-
     def __init__(self, katrain, popup, config, ignore_cats):
         self.config = config
         self.ignore_cats = ignore_cats
@@ -157,3 +157,27 @@ class ConfigPopup(QuickConfigGui):
 
         self.katrain.debug_level = self.config["debug"]["level"]
         self.katrain.update_state(redraw_board=True)
+
+
+class ConfigAIPopup(QuickConfigGui):
+    def __init__(self, katrain, popup, ai_modes, **kwargs):
+        self.settings = self.katrain.ai_settings
+        super().__init__(katrain, popup, self.settings, **kwargs)
+        self.ai_modes = ai_modes
+        Clock.schedule_once(self._build, 0)
+
+    def _build(self):
+        for mode in self.ai_modes:
+            mode_settings = self.settings[mode]
+            column = GridLayout(rows=2 + len(mode_settings), columns=2, size_hint=(0.5, 1))
+            column.add_widget(ScaledLightLabel(text=f"Settings for AI {mode}", bold=True))
+            for k, v in mode_settings.items():
+                column.add_widget(ScaledLightLabel(text=f"{k}"))
+                column.add_widget(ConfigPopup.type_to_widget_class(v)(text=str(v), input_property=f"{mode}/{k}"))
+
+                column.add_widget(Label(text=f"Settings for AI {mode}", bold=True))
+
+            self.add_widget(column)
+
+    def on_submit(self):
+        self.popup.dismiss()

katrain.kv

@@ -31,14 +31,14 @@
             radius: root.radius
 
 <StyledSpinnerOption@SpinnerOption>:
-    font_size: self.size[1] * 0.4
+    font_size: self.size[1] * 0.33
     background_color: BUTTON_COLOR
     background_normal: ''
     color: WHITE
 
 <StyledSpinner>:
     text: self.values[0] if self.values else ''
-    font_size: self.size[1] * 0.4
+    font_size: self.size[1] * 0.35
     sync_height: True
     background_color: [*[c*255/88 for c in BUTTON_COLOR[:3]], 1] # compensate for texture
     option_cls: 'StyledSpinnerOption'
@@ -449,9 +449,9 @@
                     on_selection: root.katrain.update_state()
                 StyledSpinner:
                     id: B_AI_mode
+                    values: ['Default']
                     size_hint: 0.3, 1
-                    values: AI_MODES
+                    on_text: if B_player_mode.children: B_player_mode.children[0].trigger_action(duration=0)
-                    on_text: B_player_mode.children[0].trigger_action(duration=0)
                 Label:
                     size_hint: None,1
                     width: 3
@@ -467,14 +467,24 @@
                 StyledSpinner:
                     id: W_AI_mode
                     size_hint: 0.3, 1
-                    values: AI_MODES
+                    values: ['Default']
-                    on_text: W_player_mode.children[0].trigger_action(duration=0)
+                    on_text: if W_player_mode.children: W_player_mode.children[0].trigger_action(duration=0)
+            BoxLayout:
+                size_hint: 1,0.05
+                padding: 1
+                spacing: 1
+                Label:
+                    size_hint: 0.55,1
+                StyledButton:
+                    text: 'Configure AIs'
+                    on_press: root.configure_ais()
+                    size_hint: 0.45,1
                 Label:
                     size_hint: None,1
                     width: 3
             LargeLabel:
                 text: ''
-                size_hint: 1,0.3
+                size_hint: 1,0.2
     CensorableLabel:
         id: score_change
         size_hint: 1, 0.0225

katrain.py

@@ -36,7 +36,10 @@ class KaTrainGui(BoxLayout):
         self.logger = lambda message, level=OUTPUT_INFO: self.log(message, level)
 
         self._load_config()
+
         self.debug_level = self.config("debug/level", OUTPUT_INFO)
+        self.ai_settings = self.config("ai")
+        self.controls.ai_mode_groups["W"].values = self.controls.ai_mode_groups["B"].values = self.ai_settings.keys()
         self.message_queue = Queue()
 
         self._keyboard = Window.request_keyboard(None, self, "")
@@ -91,8 +94,8 @@ class KaTrainGui(BoxLayout):
             self.game.analyze_undo(cn, self.config("trainer"))  # not via message loop
         if (
             cn.analysis_ready
-            and "ai" in self.controls.player_mode(cn.next_player)
+            and "ai" in self.controls.player_mode(cn.next_player).lower()
-            and not "pause" in self.controls.ai_mode(cn.next_player)
+            and not "pause" in self.controls.ai_mode(cn.next_player).lower()
             and not cn.children
             and not self.game.ended
             and not (auto_undo and cn.auto_undo is None)
@@ -141,7 +144,10 @@ class KaTrainGui(BoxLayout):
 
     def _do_ai_move(self, node=None):
         if node is None or self.game.current_node == node:
-            ai_move(self.game, self.controls.ai_mode(self.game.current_node.next_player), self.config("ai"))
+            mode = self.controls.ai_mode(self.game.current_node.next_player)
+            settings = self.ai_settings[mode]
+            if settings:
+                ai_move(self.game, mode, settings)
 
     def _do_undo(self, n_times=1):
         self.game.undo(n_times)

selfplay.py

@@ -37,15 +37,15 @@ class AI:
     }
     NUM_THREADS = 8
     DEFAULT_SETTINGS = {
-        "balance_target_score": 2,
+        "target_score": 2,
-        "balance_random_loss": 1,
+        "random_loss": 1,
-        "balance_max_loss": 5,
+        "max_loss": 5,
-        "balance_min_visits": 20,
+        "min_visits": 20,
         "noise_strength": 0.8,
         "pick_n": 10,
         "pick_frac": 0.2,
         "local_stddev": 10,
-        "influence_weight": 0.1,
+        "line_weight": 0.1,
         "pick_override": 0.95,
     }
     IGNORE_SETTINGS_IN_TAG = {"threads", "enable_ownership", "katago"}  # katago for switching from/to bs version
@@ -133,8 +133,8 @@ test_ais = [
     AI("P+Pick", {"pick_frac": 0.5, "pick_n": 0}),
     AI("P+Influence", {"pick_frac": 0.2, "pick_n": 20}),
     AI("P+Territory", {"pick_frac": 0.2, "pick_n": 20}),
-    AI("P+Influence", {"pick_frac": 0.33, "influence_weight": 0.05}),
+    AI("P+Influence", {"pick_frac": 0.33, "line_weight": 0.05}),
-    AI("P+Territory", {"pick_frac": 0.33, "influence_weight": 0.05}),
+    AI("P+Territory", {"pick_frac": 0.33, "line_weight": 0.05}),
     AI("P+Pick", {"pick_frac": 0.0, "pick_n": 1}),
     AI("P+Tenuki", {"local_stddev": 20}),
     AI("P+Tenuki", {"local_stddev": 10}),

start_bot

@@ -1,3 +1,6 @@
 GREETING="Hello, welcome to an experimental version of KaTrain AIs - These are based on weakened policy nets of KataGo. Current mode is: Play an influential style."
-MAXGAMES=3
+GREETING="Hello, welcome to an experimental version of KaTrain AIs - These are based on weakened policy nets of KataGo. Current mode is: Attach to ALL the things."
-gtp2ogs --apikey $(cat my/apikey) --username katrain-dev --greeting "$GREETING" --debug --ogspv katago --noclock --maxconnectedgames $MAXGAMES --persist --minrank 15k --noautohandicap --maxhandicap 0 --fakerank 3k --boardsizes 9,13,19 --komis automatic,6.5 -- python ai2gtp.py
+GREETING="Hello, welcome to an experimental version of KaTrain AIs - These are based on weakened policy nets of KataGo. Current mode is: Balanced style."
+BYEMSG="Thank you for playing. If you have any feedback, please message my admin!"
+MAXGAMES=5
+gtp2ogs --apikey $(cat my/apikey) --username katrain-dev --greeting "$GREETING" --rankedonly --farewell "$BYEMSG" --debug --ogspv katago --noclock --speeds blitz,live --maxconnectedgames $MAXGAMES --persist --minrank 15k --noautohandicap --maxhandicap 0 --fakerank 3k --boardsizes 9,13,19 --komis automatic,6.5 -- python ai2gtp.py