Skip to content

competition-guide.md

Latest commit

 

History

History
291 lines (225 loc) · 6.66 KB

competition-guide.md

File metadata and controls

291 lines (225 loc) · 6.66 KB
Raw

Competition Guide

This guide covers everything you need to know about participating in Agent Arcade competitions, including NEAR blockchain integration and competition rules.

NEAR Integration

Prerequisites

  1. NEAR Account

  2. CLI Setup

    # Login to your NEAR account
agent-arcade wallet-cmd login

# Verify connection
agent-arcade wallet-cmd status

Smart Contract Interaction

  1. View Contract

    # Check contract status
agent-arcade contract view

# View pool balance
agent-arcade pool balance

  2. Game Registration

    # Register game configuration
agent-arcade contract register-game pong \
  --min-stake 0.1 \
  --max-stake 10 \
  --min-score 0 \
  --max-score 21

Competition Rules

Participation Requirements

  1. Eligibility

    • Valid NEAR testnet account
    • Minimum stake requirement
    • Compliant agent implementation

  2. Technical Requirements

    • Use provided training environment
    • No external data/pre-training
    • Pass automated verification

Staking System

  1. Basic Rules

    • Minimum stake: 0.1 NEAR
    • Maximum stake: 10 NEAR
    • 24-hour evaluation period
    • Performance-based rewards

  2. Reward Calculation

    reward = stake_amount * (achieved_score / target_score) * multiplier
    • Multiplier ranges: 1.0x - 3.0x
    • Score thresholds vary by game

Game-Specific Rules

  1. Pong

    • Score range: 0-21
    • Minimum success: 15 points (71% win rate)
    • Maximum reward: 3x stake
    • ALE Benchmarks:
      • Random Agent: -20.7 points
      • Human Average: 9.3 points
      • Human Expert: 14.6 points
      • DQN Baseline: 18.9 points
    • Performance Tiers:
      • Bronze (1x): 8-11 points
      • Silver (1.5x): 12-14 points
      • Gold (2x): 15-17 points
      • Platinum (3x): 18+ points

  2. Space Invaders

    • Score range: 0-1000
    • Minimum success: 500 points
    • Maximum reward: 3x stake
    • ALE Benchmarks:
      • Random Agent: 148.0 points
      • Human Average: 671.0 points
      • Human Expert: 1652.0 points
      • DQN Baseline: 1076.0 points
    • Performance Tiers:
      • Bronze (1x): 300-499 points
      • Silver (1.5x): 500-749 points
      • Gold (2x): 750-999 points
      • Platinum (3x): 1000+ points
    • Special Considerations:
      • Consistent shield usage
      • Efficient bunker management
      • Strategic enemy prioritization

  3. River Raid

    • Score range: 0-15000
    • Minimum success: 10000 points
    • Maximum reward: 3x stake
    • ALE Benchmarks:
      • Random Agent: 1338.5 points
      • Human Average: 5981.9 points
      • Human Expert: 13513.3 points
      • DQN Baseline: 8627.5 points
    • Performance Tiers:
      • Bronze (1x): 5000-7499 points
      • Silver (1.5x): 7500-9999 points
      • Gold (2x): 10000-12499 points
      • Platinum (3x): 12500+ points
    • Special Considerations:
      • Fuel management efficiency
      • Bridge navigation success rate
      • Enemy avoidance patterns

Performance Metrics

  1. Core Metrics

    • Episode Score: Raw game score
    • Success Rate: % of episodes above minimum success threshold
    • Stability: Standard deviation across episodes
    • Completion Rate: % of episodes reaching target length

  2. Advanced Metrics

    • Action Efficiency: Ratio of successful to total actions
    • Resource Management: Game-specific resource usage
    • Strategic Patterns: Recognized gameplay patterns
    • Learning Progress: Score improvement over episodes

  3. Evaluation Requirements

    • Minimum 100 evaluation episodes
    • Fixed frame skip (4 frames)
    • No stochastic frame skipping
    • Deterministic evaluation
    • Standardized preprocessing:
      • Grayscale conversion
      • Frame stacking (4 frames)
      • 84x84 resolution
      • Reward clipping [-1, 1]

Evaluation Process

  1. Initial Evaluation

    # Run comprehensive evaluation
agent-arcade evaluate <game> <model-path> --episodes 100 --record

  2. Performance Verification

    • Multiple evaluation runs
    • Statistical significance checks
    • Video recording for verification
    • Automated pattern analysis

  3. Submission Requirements

    • Model checksum verification
    • Environment version matching
    • Configuration validation
    • Performance consistency check

Participation Workflow

1. Training Phase

# Train your agent
agent-arcade train pong --config models/pong/config.yaml

# Evaluate performance
agent-arcade evaluate pong models/pong/final_model.zip --episodes 100

2. Staking Phase

# Place stake
agent-arcade stake place pong \
  --model models/pong/final_model.zip \
  --amount 1 \
  --target-score 18

# Monitor stake status
agent-arcade stake view

3. Evaluation Phase

# First, run an evaluation to generate a verification token
agent-arcade evaluate pong --model models/pong/baseline/final_model.zip --episodes 50

# Submit verified score
agent-arcade stake submit pong 18

# View leaderboard
agent-arcade leaderboard top pong

The evaluation process generates a cryptographically signed verification token that proves your score was legitimately achieved. This security measure prevents arbitrary score submissions and ensures fair competition.

Reward Distribution

Calculation Example

  1. Basic Scenario

    • Stake: 1 NEAR
    • Target: 18 points
    • Achieved: 15 points
    • Reward: 1 (15/18) 1.5 = 1.25 NEAR

  2. Maximum Reward

    • Stake: 10 NEAR
    • Perfect score
    • 3x multiplier
    • Reward: 30 NEAR

Claiming Rewards

# View available rewards
agent-arcade rewards view

# Claim rewards
agent-arcade rewards claim

Best Practices

  1. Risk Management

    • Start with small stakes
    • Test thoroughly before staking
    • Monitor performance regularly

  2. Performance Optimization

    • Use evaluation metrics
    • Record validation episodes
    • Document training process

  3. Fair Play

    • Follow provided guidelines
    • Report bugs responsibly
    • Maintain code integrity

Troubleshooting

Common Issues

  1. Stake Placement

    • Insufficient funds
    • Invalid target score
    • Model validation errors

  2. Score Submission

    • Network connectivity
    • Transaction failures
    • Validation timeouts

  3. Reward Claims

    • Pending evaluations
    • Contract errors
    • Account issues

Support Channels

  1. Technical Support

    • GitHub Issues
    • Discord Community
    • Documentation

  2. Contract Issues

    • NEAR Explorer
    • Transaction logs
    • Support tickets