Spaces:
Sleeping
Sleeping
File size: 12,816 Bytes
a099b30 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 | """
Comparative Agent Arena
Run multiple agents simultaneously to compare performance.
Supports:
- Rule-based agent
- LLM agent (makes live API calls)
- Random agent
"""
import os
import json
import textwrap
import time
import random
import asyncio
from typing import Dict, List, Any, Optional
from dataclasses import dataclass, field
from datetime import datetime
from traffic_control.environment import TrafficControlEnvironment
from traffic_control.models import TrafficAction, TrafficObservation
from traffic_control.tasks import grade, GradeResult
try:
from openai import OpenAI
_HAS_OPENAI = True
except ImportError:
_HAS_OPENAI = False
@dataclass
class AgentResult:
"""Result for a single agent run."""
agent_name: str
agent_type: str
task_id: str
episode_id: str
steps: int = 0
total_reward: float = 0.0
score: float = 0.0
metrics: Dict[str, Any] = field(default_factory=dict)
grade_result: Optional[GradeResult] = None
decision_times: List[float] = field(default_factory=list)
@property
def avg_decision_time_ms(self) -> float:
if self.decision_times:
return sum(self.decision_times) / len(self.decision_times) * 1000
return 0.0
class RuleBasedAgent:
"""Rule-based traffic controller (optimized for high scores)."""
def __init__(self, name: str = "RuleBased"):
self.name = name
def decide(self, obs: TrafficObservation) -> int:
"""Return light phase based on rules."""
em_q = obs.emergency_queue
em_u = obs.emergency_urgency
q = obs.queue_lengths
current = obs.current_phase
time_in = obs.time_in_phase
# Emergency prioritization
ns_em_urgency = em_u[0] + em_u[1] + em_q[0] * 2 + em_q[1] * 2
ew_em_urgency = em_u[2] + em_u[3] + em_q[2] * 2 + em_q[3] * 2
if ns_em_urgency > 0 or ew_em_urgency > 0:
return 0 if ns_em_urgency >= ew_em_urgency else 1
# Queue-based switching
ns_total = q[0] + q[1]
ew_total = q[2] + q[3]
min_phase_time = min(3 + max(ns_total, ew_total) // 5, 8)
if current == 0 and time_in < min_phase_time and ns_total > 0:
return 0
if current == 1 and time_in < min_phase_time and ew_total > 0:
return 1
if ns_total >= ew_total + 2:
return 0
elif ew_total >= ns_total + 2:
return 1
else:
return current if current in (0, 1) else 0
class RandomAgent:
"""Random traffic controller for baseline comparison."""
def __init__(self, name: str = "Random", seed: int = 42):
self.name = name
self.rng = random.Random(seed)
def decide(self, obs: TrafficObservation) -> int:
"""Return random light phase."""
return self.rng.choice([0, 1, 2])
class RoundRobinAgent:
"""Simple round-robin controller."""
def __init__(self, name: str = "RoundRobin", switch_interval: int = 5):
self.name = name
self.switch_interval = switch_interval
self.step_count = 0
def decide(self, obs: TrafficObservation) -> int:
"""Alternate between phases."""
self.step_count += 1
phase_index = (self.step_count // self.switch_interval) % 2
return int(phase_index)
# LLM System Prompt for arena
LLM_SYSTEM_PROMPT = textwrap.dedent("""
You are an Autonomous Traffic Control AI managing a 4-way intersection.
PHASES:
0 = North-South Green (N/S vehicles may pass)
1 = East-West Green (E/W vehicles may pass)
2 = All Red (no vehicles pass)
DECISION RULES (apply in order):
1. EMERGENCY CHECK: If emergency vehicles are waiting, prioritize them.
2. MINIMUM PHASE TIME: Stay in current phase at least 3 steps if traffic present.
3. QUEUE BALANCE: Switch to direction with significantly more traffic.
OUTPUT: Reply with exactly one JSON object: {"light_phase": <0, 1, or 2>}
""")
class LLM_Agent:
"""LLM-powered agent that makes dynamic API calls."""
def __init__(
self,
name: str = "LLM-Agent",
api_base_url: Optional[str] = None,
api_key: Optional[str] = None,
model_name: str = "gpt-4.1-mini",
):
self.name = name
self.model_name = model_name
self.api_calls_made = 0
# Initialize OpenAI client if credentials available
if _HAS_OPENAI and (api_base_url or api_key):
self.client = OpenAI(
base_url=api_base_url or "https://api.openai.com/v1",
api_key=api_key or "dummy-key",
)
else:
self.client = None
def _build_prompt(self, obs: TrafficObservation) -> str:
"""Build the user prompt from observation."""
return (
f"Current phase: {obs.current_phase} (0=NS Green, 1=EW Green, 2=All Red)\n"
f"Time in phase: {obs.time_in_phase} steps\n"
f"\n"
f"Queue lengths (N, S, E, W): {obs.queue_lengths}\n"
f"Emergency queues (N, S, E, W): {obs.emergency_queue}\n"
f"Emergency urgency (N, S, E, W): {obs.emergency_urgency}\n"
f"\n"
f"What light phase should be set? Respond with JSON: {{\"light_phase\": 0, 1, or 2}}"
)
def decide(self, obs: TrafficObservation) -> int:
"""Make LLM API call to get decision."""
if not self.client:
# Fallback to rule-based if no client
return self._rule_fallback(obs)
try:
resp = self.client.chat.completions.create(
model=self.model_name,
messages=[
{"role": "system", "content": LLM_SYSTEM_PROMPT},
{"role": "user", "content": self._build_prompt(obs)},
],
temperature=0.0,
max_tokens=32,
stream=False,
)
self.api_calls_made += 1
content = resp.choices[0].message.content.strip()
# Parse JSON response
try:
data = json.loads(content)
phase = int(data.get("light_phase", 0))
return max(0, min(2, phase)) # Clamp to valid range
except (json.JSONDecodeError, ValueError, KeyError):
# Fallback if parsing fails
return self._rule_fallback(obs)
except Exception as e:
# Fallback on API error
print(f"[LLM Agent] API error: {e}, using fallback")
return self._rule_fallback(obs)
def _rule_fallback(self, obs: TrafficObservation) -> int:
"""Rule-based fallback when LLM fails."""
em_q = obs.emergency_queue
em_u = obs.emergency_urgency
q = obs.queue_lengths
current = obs.current_phase
# Emergency prioritization
ns_em = em_u[0] + em_u[1] + em_q[0] + em_q[1]
ew_em = em_u[2] + em_u[3] + em_q[2] + em_q[3]
if ns_em > 0 or ew_em > 0:
return 0 if ns_em >= ew_em else 1
# Queue-based
ns_total = q[0] + q[1]
ew_total = q[2] + q[3]
if ns_total > ew_total:
return 0
elif ew_total > ns_total:
return 1
else:
return current if current in (0, 1) else 0
class Arena:
"""Run multiple agents and compare results."""
def __init__(self):
self.results: List[AgentResult] = []
# Get LLM credentials from env (for arena LLM agent)
api_base = os.environ.get("API_BASE_URL")
api_key = os.environ.get("API_KEY")
model = os.environ.get("MODEL_NAME", "gpt-4.1-mini")
self.agents = {
"llm": LLM_Agent("Dynamic LLM", api_base, api_key, model),
"rule_based": RuleBasedAgent("Smart Rule-Based"),
"random": RandomAgent("Random Baseline"),
"round_robin": RoundRobinAgent("Round Robin"),
}
async def run_agent(
self,
agent_type: str,
task_id: str,
max_steps: int = 300,
seed: int = 42,
) -> AgentResult:
"""Run a single agent episode."""
env = TrafficControlEnvironment(task_id=task_id)
agent = self.agents.get(agent_type, self.agents["rule_based"])
obs = env.reset(seed=seed)
episode_id = env._episode_id
result = AgentResult(
agent_name=agent.name,
agent_type=agent_type,
task_id=task_id,
episode_id=episode_id,
)
for step in range(max_steps):
import time
start_time = time.time()
action_id = agent.decide(obs)
action = TrafficAction(light_phase=action_id)
decision_time = time.time() - start_time
result.decision_times.append(decision_time)
obs = env.step(action)
result.steps = step + 1
result.total_reward += obs.reward or 0.0
if obs.done:
break
# Grade the result
state = env.state
result.grade_result = grade(
task_id,
total_vehicles_passed=state.total_vehicles_passed,
total_emergency_passed=state.total_emergency_passed,
total_waiting_time=state.total_waiting_time,
total_collisions=state.total_collisions,
total_emergency_delay=state.total_emergency_delay,
total_phase_changes=state.total_phase_changes,
step_count=result.steps,
)
result.score = result.grade_result.score
result.metrics = result.grade_result.metrics
return result
async def run_comparison(
self,
task_id: str,
agents: Optional[List[str]] = None,
runs_per_agent: int = 1,
) -> Dict[str, Any]:
"""Run multiple agents and compare."""
agents_to_run = agents or list(self.agents.keys())
all_results = []
for agent_type in agents_to_run:
for run in range(runs_per_agent):
seed = 42 + run
result = await self.run_agent(agent_type, task_id, seed=seed)
all_results.append(result)
self.results.extend(all_results)
# Aggregate results
summary = self._aggregate_results(all_results)
return summary
def _aggregate_results(self, results: List[AgentResult]) -> Dict[str, Any]:
"""Aggregate results by agent type."""
by_agent: Dict[str, List[AgentResult]] = {}
for r in results:
if r.agent_type not in by_agent:
by_agent[r.agent_type] = []
by_agent[r.agent_type].append(r)
summary = {
"timestamp": datetime.now().isoformat(),
"total_runs": len(results),
"agents": {},
"winner": None,
}
best_score = -1
best_agent = None
for agent_type, agent_results in by_agent.items():
avg_score = sum(r.score for r in agent_results) / len(agent_results)
avg_reward = sum(r.total_reward for r in agent_results) / len(agent_results)
avg_steps = sum(r.steps for r in agent_results) / len(agent_results)
avg_time = sum(r.avg_decision_time_ms for r in agent_results) / len(agent_results)
summary["agents"][agent_type] = {
"name": agent_results[0].agent_name,
"runs": len(agent_results),
"avg_score": round(avg_score, 4),
"avg_total_reward": round(avg_reward, 2),
"avg_steps": round(avg_steps, 1),
"avg_decision_time_ms": round(avg_time, 2),
"best_run": max(agent_results, key=lambda r: r.score).episode_id,
}
if avg_score > best_score:
best_score = avg_score
best_agent = agent_type
summary["winner"] = best_agent
summary["all_runs"] = [
{
"agent": r.agent_type,
"episode_id": r.episode_id,
"score": r.score,
"reward": round(r.total_reward, 2),
"steps": r.steps,
}
for r in results
]
return summary
# Global arena instance
_arena = Arena()
def get_arena() -> Arena:
"""Get the global arena instance."""
return _arena
|