Multi map size training

experiments/gym-microrts-static-files/league.csv

@@ -1,14 +1,13 @@
-name,mu,sigma,trueskill
-coacAI,37.01208300442514,1.2057837814702337,33.39473166001444
-workerRushAI,32.177999126535994,1.0158657259647073,29.13040194864187
-droplet,32.046013299709365,1.0151344815092278,29.000609855181683
-mixedBot,31.485630788984253,1.0538188458547628,28.324174251419965
-izanagi,30.251048792700562,1.048141521227313,27.10662422901862
-tiamat,27.8300726697548,1.0437640107384225,24.69878063753953
-lightRushAI,26.430335172946624,1.0158591694179326,23.382757664692825
-rojo,25.174781972400247,1.0053993662036687,22.15858387378924
-guidedRojoA3N,23.137975074471896,0.9950992581462096,20.152677300033268
-naiveMCTSAI,20.932758765019557,0.9990214832647916,17.93569431522518
-randomBiasedAI,16.41033348962392,1.1908998345256558,12.837633986046953
-passiveAI,6.315587903310498,2.2292417057662304,-0.3721372139881929
-randomAI,5.9383357094823985,2.1599970738074847,-0.5416555119400552
+name,mu,sigma,trueskill
+coacAI,36.923289688405525,1.34088485677891,32.90063511806879
+workerRushAI,36.03709563706682,1.1589384846214408,32.560280183202494
+droplet,31.704029704725823,0.9728982037858849,28.785335093368168
+izanagi,32.137680892518816,1.1176715403719928,28.784666271402838
+mixedBot,30.10138243037574,1.0998952344279804,26.801696727091798
+tiamat,29.000843603866457,0.9989029800417669,26.004134663741155
+rojo,25.51046196801082,1.0193126068509855,22.452524147457865
+lightRushAI,25.200606816904337,1.1366932359941933,21.790527108921758
+naiveMCTSAI,23.1505040452353,1.12121816264434,19.786849557302283
+randomBiasedAI,11.418337600522493,1.4736936614311595,6.997256616229015
+randomAI,6.2504777818057065,1.836776105751937,0.740149464549896
+passiveAI,5.875968222282016,1.9933271395347394,-0.10401319632220218

experiments/gym-microrts-static-files/tmp/.gitignore

@@ -0,0 +1 @@
+*.db

experiments/league.py

@@ -58,7 +58,7 @@ def parse_args():
         help='the highest sigma of the trueskill evaluation')
     parser.add_argument('--output-path', type=str, default=f"league.temp.csv",
         help='the output path of the leaderboard csv')
-    parser.add_argument('--model-type', type=str, default=f"ppo_gridnet_large", choices=["ppo_gridnet_large", "ppo_gridnet"],
+    parser.add_argument('--model-type', type=str, default=f"ppo_gridnet", choices=["ppo_gridnet_large", "ppo_gridnet_variable_mapsizes", "ppo_gridnet"],
         help='the output path of the leaderboard csv')
     parser.add_argument('--maps', nargs='+', default=["maps/16x16/basesWorkers16x16A.xml"],
         help="the maps to do trueskill evaluations")
@@ -86,6 +86,10 @@ def parse_args():
 if args.model_type == "ppo_gridnet_large":
     from ppo_gridnet_large import Agent, MicroRTSStatsRecorder
 
+    from gym_microrts.envs.vec_env import MicroRTSBotVecEnv, MicroRTSGridModeVecEnv
+elif args.model_type == "ppo_gridnet_variable_mapsizes":
+    from ppo_gridnet_variable_mapsizes import Agent, MicroRTSStatsRecorder
+
     from gym_microrts.envs.vec_env import MicroRTSBotVecEnv, MicroRTSGridModeVecEnv
 else:
     from ppo_gridnet import Agent, MicroRTSStatsRecorder
@@ -231,7 +235,6 @@ def run(self, num_matches=7):
 
     def run_m0(self, num_matches):
         results = []
-        16 * 16
         next_obs = torch.Tensor(self.envs.reset()).to(self.device)
         while True:
             # self.envs.render()
@@ -256,7 +259,6 @@ def run_m0(self, num_matches):
 
     def run_m1(self, num_matches):
         results = []
-        16 * 16
         next_obs = torch.Tensor(self.envs.reset()).to(self.device)
         while True:
             # self.envs.render()

experiments/ppo_gridnet.py

@@ -356,17 +356,17 @@ def on_evaluation_done(self, future):
 
         eval_executor = ThreadPoolExecutor(max_workers=args.max_eval_workers, thread_name_prefix="league-eval-")
 
-    agent = Agent(envs).to(device)
+    mapsize = envs.height * envs.width
+    agent = Agent(envs, mapsize=mapsize).to(device)
     optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)
     if args.anneal_lr:
         # https://github.com/openai/baselines/blob/ea25b9e8b234e6ee1bca43083f8f3cf974143998/baselines/ppo2/defaults.py#L20
         lr = lambda f: f * args.learning_rate
 
     # ALGO Logic: Storage for epoch data
-    mapsize = 16 * 16
+
     action_space_shape = (mapsize, len(envs.action_plane_space.nvec))
     invalid_action_shape = (mapsize, envs.action_plane_space.nvec.sum())
-
     obs = torch.zeros((args.num_steps, args.num_envs) + envs.observation_space.shape).to(device)
     actions = torch.zeros((args.num_steps, args.num_envs) + action_space_shape).to(device)
     logprobs = torch.zeros((args.num_steps, args.num_envs)).to(device)