Osrs 5.0 pr

build.sh

@@ -87,11 +87,14 @@ fi
 PLATFORM="$(uname -s)"
 if [ "$PLATFORM" = "Linux" ]; then
     RAYLIB_NAME='raylib-5.5_linux_amd64'
+    OMP_FLAGS=(-fopenmp)
     OMP_LIB=-lomp5
     SANITIZE_FLAGS=(-fsanitize=address,undefined,bounds,pointer-overflow,leak -fno-omit-frame-pointer)
     STANDALONE_LDFLAGS=(-lGL)
 else
     RAYLIB_NAME='raylib-5.5_macos'
+    OMP_PREFIX="$(brew --prefix libomp)"
+    OMP_FLAGS=(-Xclang -fopenmp -I"$OMP_PREFIX/include" -L"$OMP_PREFIX/lib" -lomp)
     OMP_LIB=-lomp
     SANITIZE_FLAGS=()
     STANDALONE_LDFLAGS=(-framework Cocoa -framework IOKit -framework CoreVideo -framework OpenGL)
@@ -198,10 +201,11 @@ USER_OUTPUT_NAME=${OUTPUT_NAME-}
 OUTPUT_NAME=${OUTPUT_NAME:-$ENV}
 SRC_FILE=${SRC_FILE:-$SRC_DIR/$ENV.c}
 
-# Standalone environment build
-# -mavx2 enables AVX2 intrinsics (__m256, _mm256_*) which drive.h and
-# src/pufferenv.h use directly. x86_64 only — strip if porting to ARM/Apple Silicon.
-SIMD_FLAGS=(-mavx2 -mfma)
+if [ "$(uname -m)" = "x86_64" ]; then
+    SIMD_FLAGS=(-mavx2 -mfma)
+else
+    SIMD_FLAGS=()
+fi
 if [ -n "$DEBUG" ] || [ "$MODE" = "local" ]; then
     CLANG_OPT=(-g -O0 "${CLANG_WARN[@]}" "${SANITIZE_FLAGS[@]}" "${SIMD_FLAGS[@]}")
     NVCC_OPT="-O0 -g"
@@ -219,7 +223,7 @@ if [ "$MODE" = "local" ] || [ "$MODE" = "fast" ]; then
         "${LINK_ARCHIVES[@]}"
         "${EXTRA_LDFLAGS[@]}"
         "${STANDALONE_LDFLAGS[@]}"
-        -lm -lpthread -fopenmp
+        -lm -lpthread "${OMP_FLAGS[@]}"
         -DPLATFORM_DESKTOP
         "${EXTRA_CFLAGS[@]}"
     )
@@ -297,7 +301,7 @@ elif [ "$MODE" = "cpu" ]; then
         "${LINK_ARCHIVES[@]}" \
         "${EXTRA_LDFLAGS[@]}" \
         "${STANDALONE_LDFLAGS[@]}" \
-        -lm -lpthread -fopenmp \
+        -lm -lpthread "${OMP_FLAGS[@]}" \
         -o "build/cpu_${ENV}"
     echo "Built: ./build/cpu_${ENV}"
     exit 0

config/osrs_colosseum.ini

@@ -1,96 +1,90 @@
 # OSRS Fortis Colosseum encounter.
 
 [base]
-env_name = osrs_colosseum
 async = 0
 
 [env]
 start_wave = 1
-damage_reward_coeff = 0.01
-wave_clear_bonus = 0.9480448239932152
-win_bonus = 0.6761689252871952
-death_penalty_coeff = 1.0
-timeout_penalty = 0.42966922179495604
+damage_reward_coeff = 0.0058004423
+wave_clear_bonus = 0.0414097905
+win_bonus = 0.444149673
+death_penalty_coeff = 0.884283066
+timeout_penalty = 0.176490664
 farm_safe_damage_cap = 1
 farm_cap_waves = 4
 boss_damage_reward_coeff = 0.0
-boss_phase_bonus = 0.930438578582552
-argmax_gear_reward_coeff = 0.003360076095091039
-offensive_boost_reward_coeff = 0.01150336488068926
+boss_phase_bonus = 0.242805392
 loadout_profile_mode = 2
 beginner_loadout_fraction = 0.5
-step_out_forecast_obs_enabled = 1
-threat_field_obs_enabled = 1
-forecast_horizon = 4
 mask_inventory_heads = 0
-prayer_oracle_mode = 0
 late_start_state_mode = 2
 late_start_supply_fraction_per_wave = 0.07
 bis_gear_oracle_mode = 0
-invuln_mode = 0
+laser_obs_mode = 1
 episode_max_ticks_override = 0
-remove_brews = 0
 prayer_switch_fail_prob = 0.0
 player_damage_received_scale = 1.0
 damage_scale_anneal_start = 1.0
 damage_scale_anneal_ticks = 0
 classic_curriculum_mode = 1
-curriculum_num_tiers = 3
-curriculum_wave_1 = 4
-curriculum_frac_1 = 0.07367961709476761
-curriculum_wave_2 = 8
-curriculum_frac_2 = 0.03234710556986353
+curriculum_num_tiers = 4
+curriculum_wave_1 = 9
+curriculum_frac_1 = 0.0324260853
+curriculum_wave_2 = 5
+curriculum_frac_2 = 0.0615721829
 curriculum_wave_3 = 12
-curriculum_frac_3 = 0.009776253765422502
+curriculum_frac_3 = 0.16289182
+curriculum_wave_4 = 6
+curriculum_frac_4 = 0.399915189
 
 [vec]
 total_agents = 8192
 num_buffers = 8
 num_threads = 32
-action_mask_size = 452
-gpu_env = 0
+action_mask_size = 456
 
 [policy]
-hidden_size = 2048
+hidden_size = 256
 num_layers = 2
-expansion_factor = 1
 
 [train]
-total_timesteps = 143245903
-horizon = 16
-min_lr_ratio = 0.32764059472071716
-learning_rate = 0.0003724286337024778
-momentum = 0.999
-ent_coef = 1e-07
-gamma = 0.9999997409988988
-gae_lambda = 0.41662511987289663
-vtrace_rho_clip = 2.0739707238638987
-vtrace_c_clip = 3.444175403341336
-prio_alpha = 0.2617270145386202
-prio_beta0 = 0.0
-clip_coef = 0.07092591964303953
-vf_coef = 5.0
-vf_clip_coef = 0.0001
-max_grad_norm = 2.5111866624698655
-replay_ratio = 2.0
-minibatch_size = 4096
+total_timesteps = 438300800
+horizon = 32
+min_lr_ratio = 0.142049164
+learning_rate = 0.000879242609
+momentum = 0.98627466
+ent_coef = 1.41506653e-05
+gamma = 0.999796748
+gae_lambda = 0.985211074
+vtrace_rho_clip = 1.65516651
+vtrace_c_clip = 0.573781013
+prio_alpha = 0.625232637
+prio_beta0 = 0
+clip_coef = 0.109649725
+vf_coef = 4.73808718
+vf_clip_coef = 1.99999999e-06
+max_grad_norm = 6.63626146
+replay_ratio = 4.28185177
 
 [sweep]
-max_suggestion_cost = 1200
 max_runs = 400
 gpus = 1
 early_stop_quantile = 0.25
 
 [sweep.train.total_timesteps]
-distribution = log_normal
 min = 10_000_000
 max = 2_000_000_000
-scale = time
 
 [sweep.env.damage_reward_coeff]
+distribution = log_normal
+min = 0.00001
+max = 0.010
+scale = auto
+
+[sweep.env.wave_clear_bonus]
 distribution = uniform
 min = 0.0
-max = 0.010
+max = 1.0
 scale = auto
 
 [sweep.env.win_bonus]
@@ -117,26 +111,103 @@ min = 0.0
 max = 1.0
 scale = auto
 
-[sweep.env.argmax_gear_reward_coeff]
+[sweep.train.vf_coef]
+max = 24.0
+
+[sweep.train.vf_clip_coef]
+min = 0
+max = 1.0
+
+[sweep.policy.hidden_size]
+min = 64
+max = 512
+
+[sweep.policy.num_layers]
+distribution = int_uniform
+max = 4
+
+[sweep.train.horizon]
+min = 8
+max = 128
+
+[sweep.train.learning_rate]
+min = 0.00003
+max = 0.004
+
+[sweep.train.ent_coef]
+min = 0.00000001
+max = 0.002
+
+[sweep.train.gamma]
+min = 0.995
+max = 0.99997
+
+[sweep.train.gae_lambda]
+min = 0.05
+max = 0.999
+
+[sweep.train.min_lr_ratio]
 distribution = uniform
 min = 0.0
-max = 0.05
+max = 0.9
 scale = auto
 
-[sweep.env.offensive_boost_reward_coeff]
+[sweep.train.clip_coef]
+min = 0.001
+max = 0.12
+
+[sweep.train.max_grad_norm]
+min = 0.3
+max = 8.0
+
+[sweep.train.replay_ratio]
+min = 1.5
+max = 6.0
+
+[sweep.env.curriculum_frac_1]
 distribution = uniform
 min = 0.0
-max = 0.05
+max = 0.30
 scale = auto
 
-[sweep.train.vf_coef]
+[sweep.env.curriculum_frac_2]
 distribution = uniform
-min = 0.1
-max = 8.0
+min = 0.0
+max = 0.30
 scale = auto
 
-[sweep.train.vf_clip_coef]
+[sweep.env.curriculum_frac_3]
 distribution = uniform
-min = 0.0001
-max = 1.0
+min = 0.0
+max = 0.40
+scale = auto
+
+[sweep.env.curriculum_frac_4]
+distribution = uniform
+min = 0.0
+max = 0.40
+scale = auto
+
+[sweep.env.curriculum_wave_1]
+distribution = int_uniform
+min = 1
+max = 12
+scale = auto
+
+[sweep.env.curriculum_wave_2]
+distribution = int_uniform
+min = 1
+max = 12
+scale = auto
+
+[sweep.env.curriculum_wave_3]
+distribution = int_uniform
+min = 1
+max = 12
+scale = auto
+
+[sweep.env.curriculum_wave_4]
+distribution = int_uniform
+min = 1
+max = 12
 scale = auto