League evaluation for partial observable environments

experiments/league.temp.csv

@@ -0,0 +1,9 @@
+name,mu,sigma,trueskill
+POWorkerRush,38.861939273926204,3.2121311873637186,29.225545711835046
+POLightRush,25.754744223814487,2.2179711777256155,19.100830690637643
+POHeavyRush,22.98870034666808,2.212164591404348,16.352206572455035
+models/MicroRTSGridModeVecEnv__ppo_gridnet__1__1641331302/6144.pt,21.481666441292322,2.812293923609439,13.044784670464004
+models/MicroRTSGridModeVecEnv__ppo_gridnet__1__1641330893/6144.pt,16.10799174066183,5.348608643642864,0.062165809733237154
+models/MicroRTSGridModeVecEnv__ppo_gridnet__1__1641330984/6144.pt,25.0,8.333333333333334,0.0
+models/MicroRTSGridModeVecEnv__ppo_gridnet__1__1641331138/6144.pt,25.0,8.333333333333334,0.0
+PORangedRush,8.881771566353699,3.3669435969510504,-1.2190592244994534

experiments/new_league.py

@@ -56,6 +56,8 @@ def parse_args():
         help='if toggled, the database will be updated')
     parser.add_argument('--cuda', type=lambda x: bool(strtobool(x)), default=True, nargs='?', const=True,
         help='if toggled, cuda will not be enabled by default')
+    parser.add_argument('--maps', nargs='+', default=["maps/16x16/basesWorkers16x16B.xml","maps/16x16/basesWorkers16x16C.xml","maps/16x16/basesWorkers16x16D.xml", "maps/16x16/basesWorkers16x16E.xml", "maps/16x16/basesWorkers16x16F.xml"], # [],
+        help='the maps')
     # ["randomBiasedAI","workerRushAI","lightRushAI","coacAI"]
     # default=["randomBiasedAI","workerRushAI","lightRushAI","coacAI","randomAI","passiveAI","naiveMCTSAI","mixedBot","rojo","izanagi","tiamat","droplet","guidedRojoA3N"]
     args = parser.parse_args()
@@ -66,6 +68,8 @@ def parse_args():
 dbname = "league"
 if(args.partial_obs):
     dbname = 'po_league'
+
+print(dbname)
 db = SqliteDatabase(f"{dbname}.db")
 class BaseModel(Model):
     class Meta:
@@ -95,7 +99,7 @@ class Outcome(Enum):
     LOSS = -1
 
 class Match:
-    def __init__(self, partial_obs: bool, match_up=None):
+    def __init__(self, partial_obs: bool, match_up=None, map_path="maps/16x16/basesWorkers16x16A.xml"):
         # mode 0: rl-ai vs built-in-ai
         # mode 1: rl-ai vs rl-ai
         # mode 2: built-in-ai vs built-in-ai
@@ -104,6 +108,7 @@ def __init__(self, partial_obs: bool, match_up=None):
         built_in_ais2=None
         rl_ai=None
         rl_ai2=None
+        self.map_path = map_path
 
         # determine mode
         rl_ais = []
@@ -143,14 +148,15 @@ def __init__(self, partial_obs: bool, match_up=None):
         self.device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu")
         max_steps = 5000
         if mode == 0:
+            print(partial_obs)
             self.envs = MicroRTSGridModeVecEnv(
                 num_bot_envs=len(built_in_ais),
                 num_selfplay_envs=0,
                 partial_obs=partial_obs,
                 max_steps=max_steps,
                 render_theme=2,
                 ai2s=built_in_ais,
-                map_paths=["maps/16x16/basesWorkers16x16A.xml"],
+                map_paths=[map_path],
                 reward_weight=np.array([10.0, 1.0, 1.0, 0.2, 1.0, 4.0]),
             )
             self.agent = Agent(self.envs).to(self.device)
@@ -163,7 +169,7 @@ def __init__(self, partial_obs: bool, match_up=None):
                 partial_obs=partial_obs,
                 max_steps=max_steps,
                 render_theme=2,
-                map_paths=["maps/16x16/basesWorkers16x16A.xml"],
+                map_paths=[map_path],
                 reward_weight=np.array([10.0, 1.0, 1.0, 0.2, 1.0, 4.0]),
             )
             self.agent = Agent(self.envs).to(self.device)
@@ -178,7 +184,7 @@ def __init__(self, partial_obs: bool, match_up=None):
                 ai2s=built_in_ais2,
                 max_steps=max_steps,
                 render_theme=2,
-                map_paths=["maps/16x16/basesWorkers16x16.xml"],
+                map_paths=[map_path],
                 reward_weight=np.array([10.0, 1.0, 1.0, 0.2, 1.0, 4.0])
             )
         self.envs = MicroRTSStatsRecorder(self.envs)
@@ -339,41 +345,43 @@ def get_leaderboard_existing_ais(existing_ai_names):
                 if idx == 0:
                     match_up = list(reversed(match_up))
 
-                m = Match(args.partial_obs, match_up)
-                challenger = AI.get_or_none(name=m.p0)
-                defender = AI.get_or_none(name=m.p1)
-
-                r = m.run(args.num_matches // 2)
-                for item in r:
-                    drawn = False
-                    if item == Outcome.WIN.value:
-                        winner = challenger
-                        loser = defender
-                    elif item == Outcome.DRAW.value:
-                        drawn = True
-                    else:
-                        winner = defender
-                        loser = challenger
-
-                    print(f"{winner.name} {'draws' if drawn else 'wins'} {loser.name}")
-
-                    winner_rating, loser_rating = rate_1vs1(
-                        Rating(winner.mu, winner.sigma),
-                        Rating(loser.mu, loser.sigma),
-                        drawn=drawn)
-
-                    winner.mu, winner.sigma = winner_rating.mu, winner_rating.sigma
-                    loser.mu, loser.sigma = loser_rating.mu, loser_rating.sigma
-                    winner.save()
-                    loser.save()
+
+                for match in range(args.num_matches // 2):
+                    m = Match(args.partial_obs, match_up, args.maps[match])
+                    challenger = AI.get_or_none(name=m.p0)
+                    defender = AI.get_or_none(name=m.p1)
 
-                    MatchHistory(
-                        challenger=challenger,
-                        defender=defender,
-                        win=int(item == 1),
-                        draw=int(item == 0),
-                        loss=int(item == -1),
-                    ).save()
+                    r = m.run(1)
+                    for item in r:
+                        drawn = False
+                        if item == Outcome.WIN.value:
+                            winner = challenger
+                            loser = defender
+                        elif item == Outcome.DRAW.value:
+                            drawn = True
+                        else:
+                            winner = defender
+                            loser = challenger
+
+                        print(f"{winner.name} {'draws' if drawn else 'wins'} {loser.name}")
+
+                        winner_rating, loser_rating = rate_1vs1(
+                            Rating(winner.mu, winner.sigma),
+                            Rating(loser.mu, loser.sigma),
+                            drawn=drawn)
+
+                        winner.mu, winner.sigma = winner_rating.mu, winner_rating.sigma
+                        loser.mu, loser.sigma = loser_rating.mu, loser_rating.sigma
+                        winner.save()
+                        loser.save()
+
+                        MatchHistory(
+                            challenger=challenger,
+                            defender=defender,
+                            win=int(item == 1),
+                            draw=int(item == 0),
+                            loss=int(item == -1),
+                        ).save()
         get_leaderboard().to_csv(f"{dbname}.csv", index=False)
 
     # case 2: new AIs
@@ -447,7 +455,7 @@ def binary_search(leaderboard, low, high, ai, n=5):
             ai = AI.get(name=new_ai_name)
             binary_search(leaderboard, 0, len(leaderboard), ai.name, n=5)
 
-        get_leaderboard().to_csv(f"{dbname}.temp.csv", index=False)
+        get_leaderboard().to_csv(f"league.temp.csv", index=False)
 
     print("=======================")
     print(get_leaderboard())

experiments/po_league.csv

@@ -1,5 +1,5 @@
 name,mu,sigma,trueskill
-POWorkerRush,39.969639589352205,3.468613295987325,29.56379970139023
-POLightRush,26.237614972853905,2.4859644240612635,18.779721700670116
-POHeavyRush,21.253428741884882,2.4920725461958844,13.77721110329723
-PORangedRush,8.351719772683518,3.0734987337830626,-0.8687764286656705
+POWorkerRush,38.861939273926204,3.2121311873637186,29.225545711835046
+POLightRush,25.754744223814487,2.2179711777256155,19.100830690637643
+POHeavyRush,22.98870034666808,2.212164591404348,16.352206572455035
+PORangedRush,8.881771566353699,3.3669435969510504,-1.2190592244994534

experiments/ppo_gridnet.py

@@ -280,7 +280,7 @@ def get_value(self, x):
         + [microrts_ai.randomBiasedAI for _ in range(min(args.num_bot_envs, 2))]
         + [microrts_ai.lightRushAI for _ in range(min(args.num_bot_envs, 2))]
         + [microrts_ai.workerRushAI for _ in range(min(args.num_bot_envs, 2))],
-        map_paths=["maps/16x16/basesWorkers16x16.xml"],
+        map_paths=["maps/16x16/basesWorkers16x16A.xml","maps/16x16/basesWorkers16x16A.xml","maps/16x16/basesWorkers16x16B.xml","maps/16x16/basesWorkers16x16B.xml","maps/16x16/basesWorkers16x16C.xml","maps/16x16/basesWorkers16x16C.xml","maps/16x16/basesWorkers16x16D.xml","maps/16x16/basesWorkers16x16D.xml","maps/16x16/basesWorkers16x16E.xml","maps/16x16/basesWorkers16x16E.xml","maps/16x16/basesWorkers16x16F.xml","maps/16x16/basesWorkers16x16F.xml","maps/16x16/basesWorkers16x16G.xml","maps/16x16/basesWorkers16x16G.xml","maps/16x16/basesWorkers16x16H.xml","maps/16x16/basesWorkers16x16H.xml","maps/16x16/basesWorkers16x16I.xml","maps/16x16/basesWorkers16x16I.xml","maps/16x16/basesWorkers16x16J.xml","maps/16x16/basesWorkers16x16J.xml","maps/16x16/basesWorkers16x16K.xml","maps/16x16/basesWorkers16x16K.xml","maps/16x16/basesWorkers16x16L.xml","maps/16x16/basesWorkers16x16L.xml"],
         reward_weight=np.array([10.0, 1.0, 1.0, 0.2, 1.0, 4.0]),
     )
     envs = MicroRTSStatsRecorder(envs)
@@ -340,8 +340,8 @@ def get_value(self, x):
 
     ## EVALUATION LOGIC:
     eval_queue = []
-    trueskill_df = pd.read_csv("league.csv")
-    trueskill_step_df = pd.read_csv("league.csv")
+    trueskill_df = pd.read_csv("po_league.csv")
+    trueskill_step_df = pd.read_csv("po_league.csv")
     trueskill_step_df["type"] = trueskill_step_df["name"]
     trueskill_step_df["step"] = 0
     preset_trueskill_step_df = trueskill_step_df.copy()
@@ -476,7 +476,7 @@ def get_value(self, x):
                 torch.save(agent.state_dict(), f"models/{experiment_name}/agent.pt")
                 torch.save(agent.state_dict(), f"models/{experiment_name}/{global_step}.pt")
                 wandb.save(f"models/{experiment_name}/agent.pt", base_path=f"models/{experiment_name}", policy="now")
-                subprocess.Popen(["python", "new_league.py", "--evals", f"models/{experiment_name}/{global_step}.pt", "--update-db", "false"])
+                subprocess.Popen(["python", "new_league.py", "--evals", f"models/{experiment_name}/{global_step}.pt", "--update-db", "false", "--partial-obs", str(args.partial_obs)])
                 eval_queue += [f"models/{experiment_name}/{global_step}.pt"]
                 print(f"Evaluating models/{experiment_name}/{global_step}.pt")
 
@@ -500,6 +500,7 @@ def get_value(self, x):
                             "trueskill": league.loc[model_path]["trueskill"]
                         }
                         trueskill_df = trueskill_df.append(trueskill_data, ignore_index=True)
+                        print(trueskill_df)
                         wandb.log({"trueskill": wandb.Table(dataframe=trueskill_df)})
                         trueskill_data["type"] = "training"
                         trueskill_data["step"] = model_global_step