Update train_efficient.py: auto-download from HF Hub for job env
936fdd8 verified 6 months ago
25.6 kB
#!/usr/bin/env python3
"""
Orbit Wars — Efficient PPO Self-Play Training for Adaptive Parameter Controller.
Optimized version: loads agent module ONCE, modifies globals in-place each step.
"""

import copy
import math
import os
import random
import sys
import time
from collections import defaultdict, deque
from dataclasses import dataclass
from pathlib import Path
from typing import Any

import numpy as np
import torch
import torch.nn as nn
from torch.distributions import Normal

# ============================================================
# Import the base agent as a module-level namespace
# ============================================================
# Download submission.py from HF Hub if not available locally
import urllib.request
_SUBMISSION_PATH = '/app/submission.py'
if not os.path.exists(_SUBMISSION_PATH):
    print("Downloading submission.py from HF Hub...")
    urllib.request.urlretrieve(
        "https://huggingface.co/Builder-Neekhil/orbit-wars-agent/resolve/main/submission.py",
        _SUBMISSION_PATH
    )
    print("Downloaded.")

sys.path.insert(0, '/app')
_BASE_NS = {}
exec(open(_SUBMISSION_PATH).read(), _BASE_NS)
print("Base agent loaded successfully.")

# Also create a separate namespace for the opponent
_OPP_NS = {}
exec(open(_SUBMISSION_PATH).read(), _OPP_NS)

from kaggle_environments import make as _make_env

# ============================================================
# Feature Extraction
# ============================================================

FEATURE_DIM = 33

def extract_features(obs):
    get = obs.get if isinstance(obs, dict) else lambda k, d=None: getattr(obs, k, d)
    player = int(get("player", 0) or 0)
    step = int(get("step", 0) or 0)
    planets = get("planets") or []
    fleets = get("fleets") or []
    ang_vel = float(get("angular_velocity", 0.0) or 0.0)
    comet_ids = set(get("comet_planet_ids") or [])
    
    my_p = my_s = my_pr = en_p = en_s = en_pr = ne_p = ne_s = 0
    my_st = my_ro = en_st = 0
    en_by = defaultdict(int)
    
    for p in planets:
        _, owner, x, y, radius, ships, prod = p
        is_st = (math.hypot(x - 50, y - 50) + radius) >= 50.0
        if owner == player:
            my_p += 1; my_s += ships; my_pr += prod
            my_st += is_st; my_ro += (not is_st)
        elif owner == -1:
            ne_p += 1; ne_s += ships
        else:
            en_p += 1; en_s += ships; en_pr += prod; en_by[owner] += ships
            en_st += is_st
    
    my_fs = sum(f[6] for f in fleets if f[1] == player)
    en_fs = sum(f[6] for f in fleets if f[1] != player)
    my_fc = sum(1 for f in fleets if f[1] == player)
    en_fc = sum(1 for f in fleets if f[1] != player)
    mt = my_s + my_fs; et = en_s + en_fs; ta = mt + et + ne_s
    ne = len(en_by)
    mx_e = max(en_by.values()) if en_by else 0
    mn_e = min(en_by.values()) if en_by else 0
    nc = sum(1 for p in planets if p[0] in comet_ids)
    
    return np.array([
        step/500, min(1, step/100), max(0, (500-step)/500), float(step > 400),
        min(1, my_p/15), min(1, en_p/15), min(1, ne_p/15), min(1, my_st/10), min(1, my_ro/10),
        min(1, mt/max(1, ta)), min(1, et/max(1, ta)),
        math.log1p(mt)/10, math.log1p(et)/10, math.log1p(my_fs)/10, math.log1p(en_fs)/10,
        min(1, my_pr/max(1, my_pr+en_pr)), my_pr/30, en_pr/30,
        np.clip((mt-et)/max(1, ta), -1, 1), np.clip((my_p-en_p)/15, -1, 1), np.clip((my_pr-en_pr)/15, -1, 1),
        min(1, ne/3), float(ne >= 3), min(1, mx_e/max(1, et)), min(1, mn_e/max(1, mx_e+1)), min(1, en_fc/20),
        min(1, my_fc/20), my_fs/max(1, mt), en_fs/max(1, et),
        abs(ang_vel)*100, min(1, nc/5), min(1, len(planets)/30), ne_s/max(1, ta),
    ], dtype=np.float32)


class OpponentProfiler:
    def __init__(self):
        self.a = 0.1; self.agg = 0.5; self.exp = 0.5; self.trt = 0.5
        self.pp = 0; self.pf = 0; self.ps = 0; self.sc = 0

    def update(self, obs):
        get = obs.get if isinstance(obs, dict) else lambda k, d=None: getattr(obs, k, d)
        player = int(get("player", 0) or 0)
        planets = get("planets") or []; fleets = get("fleets") or []
        ep = sum(1 for p in planets if p[1] not in (-1, player))
        ef = sum(1 for f in fleets if f[1] != player)
        es = sum(p[5] for p in planets if p[1] not in (-1, player))
        es += sum(f[6] for f in fleets if f[1] != player)
        if self.sc > 0:
            fd = max(0, ef - self.pf)
            self.agg = (1-self.a)*self.agg + self.a*min(1, fd/5)
            pd = ep - self.pp
            self.exp = (1-self.a)*self.exp + self.a*np.clip(pd/3+0.5, 0, 1)
            efs = sum(f[6] for f in fleets if f[1] != player)
            t = 1 - min(1, efs/max(1, es)) if es > 0 else 0.5
            self.trt = (1-self.a)*self.trt + self.a*t
        self.pp = ep; self.pf = ef; self.ps = es; self.sc += 1
        return np.array([self.agg, self.exp, self.trt, min(1, self.sc/100), float(self.sc > 50)], dtype=np.float32)


# ============================================================
# Parameter Controller
# ============================================================

TUNABLE_PARAMS = {
    "HOSTILE_TARGET_VALUE_MULT": (2.05, 1.0, 3.0),
    "ELIMINATION_BONUS": (55.0, 10.0, 100.0),
    "PROACTIVE_DEFENSE_RATIO": (0.28, 0.05, 0.5),
    "FINISHING_HOSTILE_VALUE_MULT": (1.3, 0.8, 2.0),
    "WEAK_ENEMY_THRESHOLD": (110.0, 30.0, 200.0),
    "ATTACK_COST_TURN_WEIGHT": (0.50, 0.2, 0.8),
    "HOSTILE_MARGIN_BASE": (3.0, 1.0, 6.0),
    "FOUR_PLAYER_TARGET_MARGIN": (2.0, 0.0, 5.0),
    "FINISHING_HOSTILE_SEND_BONUS": (5.0, 1.0, 10.0),
    "STATIC_HOSTILE_VALUE_MULT": (1.65, 1.0, 2.5),
    "GANG_UP_VALUE_MULT": (1.4, 1.0, 2.0),
    "EXPOSED_PLANET_VALUE_MULT": (2.0, 1.0, 3.0),
    "REINFORCE_VALUE_MULT": (1.35, 0.8, 2.0),
    "DEFENSE_SHIP_VALUE": (0.55, 0.2, 1.0),
    "BEHIND_DOMINATION": (-0.20, -0.5, 0.0),
    "AHEAD_DOMINATION": (0.15, 0.0, 0.4),
    "LATE_REMAINING_TURNS": (70.0, 40.0, 100.0),
    "REAR_SEND_RATIO_TWO_PLAYER": (0.62, 0.3, 0.9),
    "COMET_VALUE_MULT": (0.65, 0.3, 1.2),
    "SNIPE_VALUE_MULT": (1.12, 0.7, 1.6),
}
PARAM_NAMES = list(TUNABLE_PARAMS.keys())
NUM_PARAMS = len(PARAM_NAMES)
INPUT_DIM = FEATURE_DIM + 5  # features + profile


class ParameterController(nn.Module):
    def __init__(self, input_dim=INPUT_DIM, hidden_size=128):
        super().__init__()
        self.shared = nn.Sequential(
            nn.Linear(input_dim, hidden_size), nn.ReLU(),
            nn.Linear(hidden_size, hidden_size), nn.ReLU(),
        )
        self.param_mean = nn.Sequential(
            nn.Linear(hidden_size, hidden_size // 2), nn.ReLU(),
            nn.Linear(hidden_size // 2, NUM_PARAMS),
        )
        self.param_log_std = nn.Parameter(torch.zeros(NUM_PARAMS))
        self.value_head = nn.Sequential(
            nn.Linear(hidden_size, hidden_size // 2), nn.ReLU(),
            nn.Linear(hidden_size // 2, 1),
        )

    def forward(self, x):
        h = self.shared(x)
        return torch.tanh(self.param_mean(h)), self.param_log_std, self.value_head(h).squeeze(-1)


def decode_params(raw):
    params = {}
    for i, name in enumerate(PARAM_NAMES):
        _, low, high = TUNABLE_PARAMS[name]
        t = (float(raw[i]) + 1.0) / 2.0
        params[name] = low + t * (high - low)
    return params


def apply_params(ns, params):
    """Apply parameter overrides to agent namespace (in-place, very fast)."""
    for name, value in params.items():
        if name in ns:
            ns[name] = value


def reset_params(ns):
    """Reset parameters to defaults."""
    for name, (default, _, _) in TUNABLE_PARAMS.items():
        if name in ns:
            ns[name] = default


# ============================================================
# Potential-based reward shaping
# ============================================================

def compute_potential(obs, player):
    get = obs.get if isinstance(obs, dict) else lambda k, d=None: getattr(obs, k, d)
    planets = get("planets") or []; fleets = get("fleets") or []
    my_p = my_s = my_pr = en_p = en_s = en_pr = 0
    for p in planets:
        _, owner, _, _, _, ships, prod = p
        if owner == player: my_p += 1; my_s += ships; my_pr += prod
        elif owner >= 0: en_p += 1; en_s += ships; en_pr += prod
    for f in fleets:
        _, owner, _, _, _, _, ships = f
        if owner == player: my_s += ships
        elif owner >= 0: en_s += ships
    eps = 1e-6; lr = math.log(10.0)
    pp = np.clip(math.log((my_p+eps)/(en_p+eps))/lr, -1, 1)
    ps = np.clip(math.log((my_s+eps)/(en_s+eps))/lr, -1, 1)
    pprod = np.clip(math.log((my_pr+eps)/(en_pr+eps))/lr, -1, 1)
    return 0.4*pp + 0.3*ps + 0.3*pprod


# ============================================================
# Efficient training loop  
# ============================================================

def run_episode_vs_random(learner_ns, seed, learner_slot=0):
    """Run episode against Kaggle's built-in random agent (very fast)."""
    from kaggle_environments.envs.orbit_wars.orbit_wars import random_agent
    
    env = _make_env("orbit_wars", configuration={"seed": seed}, debug=False)
    env.reset(num_agents=2)
    learner_ns['_agent_step'] = 0
    
    profiler = OpponentProfiler()
    states = env.step([[], []])
    learner_obs = states[learner_slot].observation
    
    features = extract_features(learner_obs)
    profile = profiler.update(learner_obs)
    initial_obs_vec = np.concatenate([features, profile])
    
    player = int(learner_obs.get("player", 0) if isinstance(learner_obs, dict) else learner_obs.player)
    prev_potential = compute_potential(learner_obs, player)
    total_shaped_reward = 0.0
    step_count = 0
    done = False
    
    while not done:
        try:
            learner_moves = learner_ns['agent'](learner_obs)
        except Exception:
            learner_moves = []
        
        opp_obs = states[1 - learner_slot].observation
        try:
            opponent_moves = random_agent(opp_obs)
        except Exception:
            opponent_moves = []
        
        if learner_slot == 0:
            actions = [learner_moves, opponent_moves]
        else:
            actions = [opponent_moves, learner_moves]
        
        states = env.step(actions)
        learner_state = states[learner_slot]
        learner_obs = learner_state.observation
        done = learner_state.status != "ACTIVE"
        
        curr_potential = compute_potential(learner_obs, player)
        step_reward = 0.99 * curr_potential - prev_potential
        prev_potential = curr_potential
        
        if done:
            raw_reward = float(learner_state.reward) if learner_state.reward else 0.0
            step_reward += raw_reward
        
        total_shaped_reward += step_reward
        step_count += 1
        profile = profiler.update(learner_obs)
    
    final_features = extract_features(learner_obs)
    final_obs_vec = np.concatenate([final_features, profile])
    final_reward = float(learner_state.reward) if learner_state.reward else 0.0
    
    return initial_obs_vec, final_obs_vec, total_shaped_reward, final_reward, step_count


def run_episode(learner_ns, opponent_ns, seed, learner_slot=0):
    """Run a full game episode. Returns (transitions, final_reward).
    
    Each transition: (features, reward, done)
    The controller makes ONE decision per episode (parameter setting for the whole game).
    This is much more efficient than per-step parameter tuning.
    """
    env = _make_env("orbit_wars", configuration={"seed": seed}, debug=False)
    env.reset(num_agents=2)
    
    # Reset step counters in both agents
    learner_ns['_agent_step'] = 0
    opponent_ns['_agent_step'] = 0
    
    profiler = OpponentProfiler()
    
    # Collect initial observation
    states = env.step([[], []])
    learner_obs = states[learner_slot].observation
    opp_obs = states[1 - learner_slot].observation
    
    # Extract initial features for controller decision
    features = extract_features(learner_obs)
    profile = profiler.update(learner_obs)
    initial_obs_vec = np.concatenate([features, profile])
    
    prev_potential = compute_potential(learner_obs, 
        int(learner_obs.get("player", 0) if isinstance(learner_obs, dict) else learner_obs.player))
    
    total_shaped_reward = 0.0
    step_count = 0
    done = False
    
    # Run the full game
    while not done:
        # Get moves from both agents
        try:
            learner_moves = learner_ns['agent'](learner_obs)
        except Exception:
            learner_moves = []
        
        try:
            opponent_moves = opponent_ns['agent'](opp_obs)
        except Exception:
            opponent_moves = []
        
        if learner_slot == 0:
            actions = [learner_moves, opponent_moves]
        else:
            actions = [opponent_moves, learner_moves]
        
        states = env.step(actions)
        learner_state = states[learner_slot]
        opp_state = states[1 - learner_slot]
        
        learner_obs = learner_state.observation
        opp_obs = opp_state.observation
        done = learner_state.status != "ACTIVE"
        
        # Shaped reward
        player = int(learner_obs.get("player", 0) if isinstance(learner_obs, dict) else learner_obs.player)
        curr_potential = compute_potential(learner_obs, player)
        step_reward = 0.99 * curr_potential - prev_potential
        prev_potential = curr_potential
        
        if done:
            raw_reward = float(learner_state.reward) if learner_state.reward else 0.0
            step_reward += raw_reward
        
        total_shaped_reward += step_reward
        step_count += 1
        
        # Update opponent profile
        profile = profiler.update(learner_obs)
    
    # Final features for the last state
    final_features = extract_features(learner_obs)
    final_obs_vec = np.concatenate([final_features, profile])
    
    final_reward = float(learner_state.reward) if learner_state.reward else 0.0
    
    return initial_obs_vec, final_obs_vec, total_shaped_reward, final_reward, step_count


def train():
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Device: {device}")
    
    # Config
    total_updates = int(os.environ.get("TOTAL_UPDATES", "500"))
    episodes_per_update = int(os.environ.get("EPISODES_PER_UPDATE", "4"))
    eval_every = int(os.environ.get("EVAL_EVERY", "25"))
    eval_games = int(os.environ.get("EVAL_GAMES", "6"))
    lr = float(os.environ.get("LR", "3e-4"))
    gamma = 0.99
    clip_coef = 0.2
    ent_coef = 0.01
    vf_coef = 0.5
    epochs = 4
    pool_size = 3
    save_dir = Path(os.environ.get("SAVE_DIR", "/app/checkpoints"))
    save_dir.mkdir(parents=True, exist_ok=True)
    random.seed(42); np.random.seed(42); torch.manual_seed(42)
    
    controller = ParameterController().to(device)
    optimizer = torch.optim.Adam(controller.parameters(), lr=lr)
    
    # Opponent pool: list of parameter snapshots (dicts of param values)
    opponent_pool = [None]  # None = baseline (no overrides)
    best_win_rate = 0.0
    seed_counter = 0
    
    # Import fast opponents
    from kaggle_environments.envs.orbit_wars.orbit_wars import random_agent
    
    # Opponent curriculum: random first, then baseline, then self-play
    def get_opponent_ns(update_idx):
        """Return opponent namespace and label based on training phase."""
        phase_fraction = update_idx / total_updates
        
        if phase_fraction < 0.2:
            # Phase 1: Train vs random (very fast ~20s/episode)
            return None, "random"
        elif phase_fraction < 0.5:
            # Phase 2: Train vs baseline (medium ~60s/episode)
            reset_params(_OPP_NS)
            return _OPP_NS, "baseline"
        else:
            # Phase 3: Train vs pool (self-play)
            opp_params = random.choice(opponent_pool)
            reset_params(_OPP_NS)
            if opp_params is not None:
                apply_params(_OPP_NS, opp_params)
            return _OPP_NS, "pool"
    
    print(f"\nTraining: {total_updates} updates × {episodes_per_update} episodes")
    print(f"Phase 1 (0-20%): vs random | Phase 2 (20-50%): vs baseline | Phase 3 (50-100%): self-play")
    print(f"Eval every {eval_every} updates, {eval_games} games\n")
    
    for update in range(total_updates):
        t0 = time.time()
        
        # Collect episodes
        obs_batch = []
        reward_batch = []
        wins = 0
        total_steps = 0
        
        for ep in range(episodes_per_update):
            seed_counter += 1
            learner_slot = (update * episodes_per_update + ep) % 2
            
            # Pick opponent based on curriculum
            opp_ns, opp_label = get_opponent_ns(update)
            
            # Get controller output for this episode
            with torch.inference_mode():
                # Use a dummy observation to get initial params
                # (we'll use the same params for the whole episode)
                dummy_obs = np.zeros(INPUT_DIM, dtype=np.float32)
                dummy_obs[0] = 0.0  # start of game
                x = torch.from_numpy(dummy_obs).unsqueeze(0).to(device)
                param_mean, log_std, value = controller(x)
                
                std = torch.exp(log_std)
                dist = Normal(param_mean.squeeze(0), std)
                action = dist.sample()
                log_prob = dist.log_prob(action).sum().item()
                value_np = value.item()
                action_np = action.cpu().numpy()
            
            # Apply learned params to learner
            params = decode_params(np.clip(action_np, -1, 1))
            reset_params(_BASE_NS)
            apply_params(_BASE_NS, params)
            
            # Run episode
            if opp_ns is None:
                # Use random agent (fast)
                init_obs, final_obs, shaped_reward, raw_reward, steps = run_episode_vs_random(
                    _BASE_NS, seed=seed_counter * 37 + 1, learner_slot=learner_slot
                )
            else:
                init_obs, final_obs, shaped_reward, raw_reward, steps = run_episode(
                    _BASE_NS, opp_ns, seed=seed_counter * 37 + 1, learner_slot=learner_slot
                )
            
            obs_batch.append((init_obs, action_np, log_prob, value_np, shaped_reward))
            reward_batch.append(raw_reward)
            if raw_reward > 0:
                wins += 1
            total_steps += steps
        
        # PPO update
        if obs_batch:
            obs_t = torch.tensor(np.stack([o[0] for o in obs_batch]), dtype=torch.float32, device=device)
            actions_t = torch.tensor(np.stack([o[1] for o in obs_batch]), dtype=torch.float32, device=device)
            old_log_probs_t = torch.tensor([o[2] for o in obs_batch], dtype=torch.float32, device=device)
            old_values_t = torch.tensor([o[3] for o in obs_batch], dtype=torch.float32, device=device)
            rewards_t = torch.tensor([o[4] for o in obs_batch], dtype=torch.float32, device=device)
            
            # Returns = rewards (single step per "episode" from controller's perspective)
            returns_t = rewards_t
            advantages_t = returns_t - old_values_t
            if advantages_t.std() > 1e-6:
                advantages_t = (advantages_t - advantages_t.mean()) / (advantages_t.std() + 1e-8)
            
            metrics = {"loss": 0, "pl": 0, "vl": 0, "ent": 0}
            n_updates = 0
            
            for _ in range(epochs):
                param_mean, log_std, values = controller(obs_t)
                std = torch.exp(log_std)
                dist = Normal(param_mean, std)
                new_log_probs = dist.log_prob(actions_t).sum(-1)
                entropy = dist.entropy().sum(-1)
                
                ratio = (new_log_probs - old_log_probs_t).exp()
                s1 = -advantages_t * ratio
                s2 = -advantages_t * torch.clamp(ratio, 1 - clip_coef, 1 + clip_coef)
                pl = torch.max(s1, s2).mean()
                vl = 0.5 * (returns_t - values).pow(2).mean()
                el = -entropy.mean()
                
                loss = pl + vf_coef * vl + ent_coef * el
                optimizer.zero_grad()
                loss.backward()
                nn.utils.clip_grad_norm_(controller.parameters(), 0.5)
                optimizer.step()
                
                metrics["loss"] += loss.item()
                metrics["pl"] += pl.item()
                metrics["vl"] += vl.item()
                metrics["ent"] += entropy.mean().item()
                n_updates += 1
            
            metrics = {k: v / max(1, n_updates) for k, v in metrics.items()}
        
        elapsed = time.time() - t0
        win_rate = wins / episodes_per_update
        avg_reward = np.mean(reward_batch) if reward_batch else 0
        
        print(f"U{update+1:4d}/{total_updates} | "
              f"WR: {win_rate:.0%} | R: {avg_reward:+.2f} | "
              f"L: {metrics.get('loss',0):.4f} PL: {metrics.get('pl',0):.4f} "
              f"VL: {metrics.get('vl',0):.4f} Ent: {metrics.get('ent',0):.3f} | "
              f"Steps: {total_steps} | {elapsed:.1f}s | vs: {opp_label}")
        
        # Evaluation and pool management
        if (update + 1) % eval_every == 0:
            print(f"\n  Evaluating vs baseline ({eval_games} games)...")
            eval_wins = 0
            
            # Get current best params from controller  
            with torch.inference_mode():
                x = torch.zeros(1, INPUT_DIM, device=device)
                pm, _, _ = controller(x)
                eval_params = decode_params(pm.squeeze(0).cpu().numpy())
            
            for g in range(eval_games):
                slot = g % 2
                reset_params(_BASE_NS); apply_params(_BASE_NS, eval_params)
                reset_params(_OPP_NS)  # opponent = baseline
                
                _, _, _, raw_r, _ = run_episode(_BASE_NS, _OPP_NS, seed=10000 + g, learner_slot=slot)
                if raw_r > 0:
                    eval_wins += 1
                print(f"    Game {g+1}: {'WIN' if raw_r > 0 else 'LOSS'} (slot={slot})")
            
            wr = eval_wins / eval_games
            print(f"  Win rate: {wr:.0%} ({eval_wins}/{eval_games})")
            
            # Add to pool if good
            if wr >= 0.45:
                if len(opponent_pool) >= pool_size:
                    opponent_pool.pop(0)
                opponent_pool.append(copy.deepcopy(eval_params))
                print(f"  ✓ Added to pool (size={len(opponent_pool)})")
            
            if wr > best_win_rate:
                best_win_rate = wr
                torch.save({
                    "controller": controller.state_dict(),
                    "params": eval_params,
                    "win_rate": wr,
                    "update": update + 1,
                }, save_dir / "best_controller.pt")
                print(f"  ★ New best: {wr:.0%}")
            print()
        
        # Checkpoint
        if (update + 1) % 100 == 0:
            torch.save({
                "controller": controller.state_dict(),
                "optimizer": optimizer.state_dict(),
                "update": update + 1,
            }, save_dir / f"ckpt_{update+1:05d}.pt")
    
    # Final save
    torch.save({
        "controller": controller.state_dict(),
        "best_win_rate": best_win_rate,
    }, save_dir / "final_controller.pt")
    
    print(f"\nDone! Best win rate: {best_win_rate:.0%}")
    print(f"Checkpoints: {save_dir}")
    
    # Push to hub
    try:
        from huggingface_hub import HfApi
        api = HfApi(token=os.environ.get("HF_TOKEN"))
        
        # Upload best checkpoint
        best_path = save_dir / "best_controller.pt"
        if best_path.exists():
            api.upload_file(
                path_or_fileobj=str(best_path),
                path_in_repo="best_controller.pt",
                repo_id="Builder-Neekhil/orbit-wars-agent",
                commit_message=f"Upload trained controller (WR: {best_win_rate:.0%})"
            )
            print(f"Uploaded best_controller.pt to HF Hub")
        
        # Generate and upload adaptive submission
        final_path = save_dir / "final_controller.pt"
        if not best_path.exists():
            best_path = final_path
        if best_path.exists():
            # Download generate_submission.py from HF Hub
            gen_script = '/app/generate_submission.py'
            if not os.path.exists(gen_script):
                urllib.request.urlretrieve(
                    "https://huggingface.co/Builder-Neekhil/orbit-wars-agent/resolve/main/generate_submission.py",
                    gen_script
                )
            sys.path.insert(0, '/app')
            from generate_submission import generate_submission
            generate_submission(
                base_agent_path=_SUBMISSION_PATH,
                checkpoint_path=str(best_path),
                output_path="/app/submission_adaptive.py",
            )
            api.upload_file(
                path_or_fileobj="/app/submission_adaptive.py",
                path_in_repo="submission_adaptive.py",
                repo_id="Builder-Neekhil/orbit-wars-agent",
                commit_message=f"Upload adaptive submission (WR: {best_win_rate:.0%})"
            )
            print("Uploaded submission_adaptive.py to HF Hub")
    except Exception as e:
        print(f"Hub upload error: {e}")


if __name__ == "__main__":
    train()