mirror of
https://github.com/GoodStartLabs/AI_Diplomacy.git
synced 2026-04-19 12:58:09 +00:00
Add benchmark improvements and game state management
Changes: - Set default num_negotiation_rounds to 3 (was 0) for better gameplay - Improved game state loading/resuming with fallback agent initialization - Added support for loading relationships from older save formats - Enhanced state update prompts with current goals and relationships - Better error handling for missing state_agents in saved games 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
AlxAI
parent
701790a37a
commit
941fe05cd3
6 changed files with 76 additions and 30 deletions
|
|
@ -999,6 +999,10 @@ class DiplomacyAgent:
|
|||
# Extract year from the phase name (e.g., "S1901M" -> "1901")
|
||||
current_year = last_phase_name[1:5] if len(last_phase_name) >= 5 else "unknown"
|
||||
|
||||
# Format current goals and relationships for the prompt
|
||||
current_goals_str = json.dumps(self.goals, indent=2) if self.goals else "[]"
|
||||
current_relationships_str = json.dumps(self.relationships, indent=2) if self.relationships else "{}"
|
||||
|
||||
prompt = prompt_template.format(
|
||||
power_name=power_name,
|
||||
current_year=current_year,
|
||||
|
|
@ -1006,6 +1010,8 @@ class DiplomacyAgent:
|
|||
board_state_str=context,
|
||||
phase_summary=last_phase_summary, # Use provided phase_summary
|
||||
other_powers=str(other_powers), # Pass as string representation
|
||||
current_goals=current_goals_str,
|
||||
current_relationships=current_relationships_str,
|
||||
)
|
||||
logger.debug(f"[{power_name}] State update prompt:\n{prompt}")
|
||||
|
||||
|
|
|
|||
|
|
@ -172,11 +172,15 @@ def load_game_state(
|
|||
game_file_name: str,
|
||||
run_config,
|
||||
resume_from_phase: Optional[str] = None,
|
||||
) -> Tuple["Game", Dict[str, "DiplomacyAgent"], "GameHistory", Optional[Any]]:
|
||||
) -> Tuple["Game", Dict[str, "DiplomacyAgent"], "GameHistory", Optional[Any], Optional[Dict[str, Dict[str, str]]]]:
|
||||
"""
|
||||
Load and fully re-hydrate the game, agents and GameHistory – including
|
||||
`orders_by_power`, `results_by_power`, `submitted_orders_by_power`,
|
||||
and per-power `phase_summaries`.
|
||||
|
||||
Returns a 5-tuple: (game, agents, game_history, run_config, saved_relationships)
|
||||
where saved_relationships is extracted from agent_relationships field (old format)
|
||||
when state_agents is not available.
|
||||
"""
|
||||
from collections import defaultdict # local to avoid new global import
|
||||
|
||||
|
|
@ -235,24 +239,37 @@ def load_game_state(
|
|||
else:
|
||||
raise ValueError(f"Invalid --models argument: expected 1 or {len(powers_order)} items, got {len(provided)}.")
|
||||
|
||||
# Extract relationships from previous phase for agent initialization
|
||||
saved_relationships = None
|
||||
if saved_game_data.get("phases"):
|
||||
last_phase_data = saved_game_data["phases"][-2] if len(saved_game_data["phases"]) > 1 else {}
|
||||
if "state_agents" not in last_phase_data:
|
||||
raise ValueError("Cannot resume: 'state_agents' key missing in last completed phase.")
|
||||
|
||||
for power_name, agent_data in last_phase_data["state_agents"].items():
|
||||
override_id = power_model_map.get(power_name)
|
||||
prompts_dir_from_config = (
|
||||
run_config.prompts_dir_map.get(power_name)
|
||||
if getattr(run_config, "prompts_dir_map", None)
|
||||
else run_config.prompts_dir
|
||||
)
|
||||
agents[power_name] = deserialize_agent(
|
||||
agent_data,
|
||||
prompts_dir=prompts_dir_from_config,
|
||||
override_model_id=override_id,
|
||||
override_max_tokens=model_max_tokens.get(power_name),
|
||||
)
|
||||
# Try to load relationships from agent_relationships field (older format)
|
||||
if "agent_relationships" in last_phase_data:
|
||||
saved_relationships = last_phase_data["agent_relationships"]
|
||||
logger.info(
|
||||
"Loaded agent_relationships from previous phase. "
|
||||
"Agents will be initialized fresh but with historical relationships."
|
||||
)
|
||||
else:
|
||||
logger.warning(
|
||||
"Cannot resume agents: 'state_agents' key missing in last completed phase. "
|
||||
"Agents will be initialized fresh (losing prior context/relationships)."
|
||||
)
|
||||
else:
|
||||
for power_name, agent_data in last_phase_data["state_agents"].items():
|
||||
override_id = power_model_map.get(power_name)
|
||||
prompts_dir_from_config = (
|
||||
run_config.prompts_dir_map.get(power_name)
|
||||
if getattr(run_config, "prompts_dir_map", None)
|
||||
else run_config.prompts_dir
|
||||
)
|
||||
agents[power_name] = deserialize_agent(
|
||||
agent_data,
|
||||
prompts_dir=prompts_dir_from_config,
|
||||
override_model_id=override_id,
|
||||
override_max_tokens=model_max_tokens.get(power_name),
|
||||
)
|
||||
|
||||
# --- Rebuild GameHistory --------------------------------------------------
|
||||
game_history = GameHistory()
|
||||
|
|
@ -297,7 +314,7 @@ def load_game_state(
|
|||
submitted[pwr].append(order_str)
|
||||
ph_obj.submitted_orders_by_power = submitted
|
||||
|
||||
return game, agents, game_history, run_config
|
||||
return game, agents, game_history, run_config, saved_relationships
|
||||
|
||||
|
||||
# ai_diplomacy/game_logic.py
|
||||
|
|
@ -306,8 +323,15 @@ async def initialize_new_game(
|
|||
game: Game,
|
||||
game_history: GameHistory,
|
||||
llm_log_file_path: str,
|
||||
saved_relationships: Optional[Dict[str, Dict[str, str]]] = None,
|
||||
) -> Dict[str, DiplomacyAgent]:
|
||||
"""Initializes agents for a new game (supports per-power prompt directories)."""
|
||||
"""
|
||||
Initializes agents for a new game (supports per-power prompt directories).
|
||||
|
||||
Args:
|
||||
saved_relationships: Optional historical relationships to restore from old game format.
|
||||
Dict mapping power_name -> {other_power: relationship_status}
|
||||
"""
|
||||
|
||||
powers_order = sorted(list(ALL_POWERS))
|
||||
|
||||
|
|
@ -357,9 +381,17 @@ async def initialize_new_game(
|
|||
try:
|
||||
client = load_model_client(model_id, prompts_dir=prompts_dir_for_power)
|
||||
client.max_tokens = model_max_tokens[power_name]
|
||||
|
||||
# Extract historical relationships if available
|
||||
initial_relationships = None
|
||||
if saved_relationships and power_name in saved_relationships:
|
||||
initial_relationships = saved_relationships[power_name]
|
||||
logger.info(f"[{power_name}] Restoring historical relationships: {initial_relationships}")
|
||||
|
||||
agent = DiplomacyAgent(
|
||||
power_name=power_name,
|
||||
client=client,
|
||||
initial_relationships=initial_relationships,
|
||||
prompts_dir=prompts_dir_for_power,
|
||||
)
|
||||
agents[power_name] = agent
|
||||
|
|
|
|||
|
|
@ -59,6 +59,8 @@ async def initialize_agent_state_ext(
|
|||
|
||||
formatted_diary = agent.format_private_diary_for_prompt()
|
||||
|
||||
# Pass agent's current relationships to provide historical context to the LLM
|
||||
# This ensures the LLM sees any pre-existing relationship state when setting initial goals
|
||||
context = build_context_prompt(
|
||||
game=game,
|
||||
board_state=board_state,
|
||||
|
|
@ -66,7 +68,7 @@ async def initialize_agent_state_ext(
|
|||
possible_orders=None, # Don't include orders for initial state setup
|
||||
game_history=game_history,
|
||||
agent_goals=None,
|
||||
agent_relationships=None,
|
||||
agent_relationships=agent.relationships, # Pass agent's relationships for LLM context
|
||||
agent_private_diary=formatted_diary,
|
||||
prompts_dir=effective_prompts_dir,
|
||||
)
|
||||
|
|
|
|||
|
|
@ -63,7 +63,8 @@ def assign_models_to_powers() -> Dict[str, str]:
|
|||
Models supported: o3-mini, o4-mini, o3, gpt-4o, gpt-4o-mini,
|
||||
claude-opus-4-20250514, claude-sonnet-4-20250514, claude-3-5-haiku-20241022, claude-3-5-sonnet-20241022, claude-3-7-sonnet-20250219
|
||||
gemini-2.0-flash, gemini-2.5-flash-preview-04-17, gemini-2.5-pro-preview-03-25,
|
||||
deepseek-chat, deepseek-reasoner
|
||||
deepseek-chat, deepseek-reasoner,
|
||||
kimi-k2-0905-preview,
|
||||
openrouter-meta-llama/llama-3.3-70b-instruct, openrouter-qwen/qwen3-235b-a22b, openrouter-microsoft/phi-4-reasoning-plus:free,
|
||||
openrouter-deepseek/deepseek-prover-v2:free, openrouter-meta-llama/llama-4-maverick:free, openrouter-nvidia/llama-3.3-nemotron-super-49b-v1:free,
|
||||
openrouter-google/gemma-3-12b-it:free, openrouter-google/gemini-2.5-flash-preview-05-20
|
||||
|
|
@ -72,13 +73,13 @@ def assign_models_to_powers() -> Dict[str, str]:
|
|||
# POWER MODELS
|
||||
|
||||
return {
|
||||
"AUSTRIA": "o4-mini",
|
||||
"ENGLAND": "o3",
|
||||
"FRANCE": "gpt-5-reasoning-alpha-2025-07-19",
|
||||
"GERMANY": "gpt-4.1",
|
||||
"ITALY": "o4-mini",
|
||||
"RUSSIA": "gpt-5-reasoning-alpha-2025-07-19",
|
||||
"TURKEY": "o4-mini",
|
||||
"AUSTRIA": "gpt-5-mini",
|
||||
"ENGLAND": "gpt-5-mini",
|
||||
"FRANCE": "gpt-5-mini",
|
||||
"GERMANY": "gpt-5-mini",
|
||||
"ITALY": "gemini-2.5-flash",
|
||||
"RUSSIA": "gemini-2.5-flash",
|
||||
"TURKEY": "gemini-2.5-flash",
|
||||
}
|
||||
|
||||
# TEST MODELS
|
||||
|
|
|
|||
|
|
@ -154,7 +154,7 @@ def _add_lm_game_flags(p: argparse.ArgumentParser) -> None:
|
|||
p.add_argument(
|
||||
"--num_negotiation_rounds",
|
||||
type=int,
|
||||
default=0,
|
||||
default=3,
|
||||
help="Number of negotiation rounds per phase.",
|
||||
)
|
||||
p.add_argument(
|
||||
|
|
|
|||
|
|
@ -100,7 +100,7 @@ def parse_arguments():
|
|||
parser.add_argument(
|
||||
"--num_negotiation_rounds",
|
||||
type=int,
|
||||
default=0,
|
||||
default=3,
|
||||
help="Number of negotiation rounds per phase.",
|
||||
)
|
||||
parser.add_argument(
|
||||
|
|
@ -311,9 +311,14 @@ async def main():
|
|||
if is_resuming:
|
||||
try:
|
||||
# When resuming, we always use the provided params (they will override the params used in the saved state)
|
||||
game, agents, game_history, _ = load_game_state(run_dir, game_file_name, run_config, args.resume_from_phase)
|
||||
game, agents, game_history, _, saved_relationships = load_game_state(run_dir, game_file_name, run_config, args.resume_from_phase)
|
||||
|
||||
logger.info(f"Successfully resumed game from phase: {game.get_current_phase()}.")
|
||||
|
||||
# If agents is empty (state_agents missing from old game format), initialize fresh agents
|
||||
if not agents:
|
||||
logger.warning("No agents loaded from game state. Initializing fresh agents for all powers.")
|
||||
agents = await initialize_new_game(run_config, game, game_history, llm_log_file_path, saved_relationships=saved_relationships)
|
||||
except (FileNotFoundError, ValueError) as e:
|
||||
logger.error(f"Could not resume game: {e}. Starting a new game instead.")
|
||||
is_resuming = False # Fallback to new game
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue