5.0 Self Play Env - Admiral

config/admiral.ini

@@ -0,0 +1,76 @@
+[base]
+env_name = admiral
+checkpoint_interval = 8
+seed = 42
+
+[vec]
+total_agents = 4096
+num_buffers = 1
+num_threads = 4
+num_policies = 2
+hist_policy_percent = 0.1
+hist_policy_hidden_size = 128
+hist_policy_num_layers = 2
+
+[selfplay]
+enabled = 1
+max_size = 100
+seed = 42
+opp_timeout_steps = 16_777_216
+eval_games = 4096
+eval_pool_size = 8
+
+[env]
+num_agents = 2 # Must match N_TEAMS; also used by the shared evaluator.
+curriculum_level = 1
+width = 2000
+height = 2000
+reward_damage_mult = 0.5
+reward_kill = 0.19897151
+penalty_hit_ally = -0.542492926
+penalty_used_volley = 0
+penalty_stationary = -0.00454199314
+
+[policy]
+hidden_size = 128
+num_layers = 2
+
+[train]
+gpus = 1
+total_timesteps = 536_870_912
+learning_rate = 0.000396882882
+anneal_lr = 0
+gamma = 0.995000005
+gae_lambda = 0.979191363
+replay_ratio = 1.95889282
+clip_coef = 0.169899464
+vf_coef = 3
+vf_clip_coef = 0.100000001
+max_grad_norm = 1.38442755
+ent_coef = 3.39512771e-05
+anneal_ent_coef = 0
+momentum = 0.997428834
+minibatch_size = 8192
+horizon = 128
+
+[sweep]
+# Self-play pool evaluation replaces this training metric as Protein's objective.
+metric = perf
+metric_distribution = linear
+goal = maximize
+max_runs = 1000
+gpus = 1
+downsample = 32
+prune_pareto = True
+
+[sweep.train.learning_rate]
+distribution = log_normal
+min = 0.0003
+max = 0.003
+scale = 0.5
+
+[sweep.train.ent_coef]
+distribution = log_normal
+min = 0.00001
+max = 0.001
+scale = auto