ai-tournament

A modular Rust crate system for evaluating AI agents via customizable tournaments, supporting sandboxed execution and flexible constraints.

Overview

This project provides tools to benchmark and evaluate AI agents in a controlled tournament setting. It supports custom games, custom tournament strategies, and execution isolation using Linux cgroups v2.

Key Features

• Pluggable Tournaments: Define your own tournament logic via the TournamentStrategy trait, or use built-in strategies like SwissTournament and SinglePlayerTournament.

• Custom Games: Any environment that implements the Game trait can be used.

• Sandboxed Agent Execution: Each agent runs in its own isolated process with:

  • Dedicated CPU cores (taskset)
  • Memory and CPU limits (via cgroups v2)
  • Action timeouts and total time budgets

• Configurable Constraints: Use the ConstraintsBuilder to define:

  • CPUs used per agent
  • Memory limits
  • Timeouts and think-time budgets

[!NOTE] Full CPU and RAM isolation requires Linux with cgroups v2 and the taskset command installed.
If not available, the evaluator can optionally fall back to time-only constraints by setting allow_uncontained = true in the configuration.

Evaluator Configuration

The Configuration struct controls how evaluation is performed. You can use Configuration::new() for defaults or customize it via builder methods.
You can also override its behavior using environment variables (EVAL_VERBOSE, EVAL_ALLOW_UNCONTAINED, etc.). See configuration.rs for details.

Usage Summary

1. Implement the Game trait for your task or environment.
2. Provide AI agents as Rust crates or compiled binaries in a specified directory.
3. Define resource constraints with the ConstraintsBuilder.
4. Choose or implement a TournamentStrategy.
5. Run the evaluator to get per-agent scores, as defined by the tournament type.

Agent Directory Structure

There are two ways to organize agent directories depending on whether you're compiling them:

If compile_agents = true (default):

Each agent should be a Rust crate with a YAML config file at the root:

agent_directory/
├── Cargo.toml
├── src/
└── config.yaml

If compile_agents = false:

Each agent subdirectory should contain a precompiled binary and a YAML config:

agent_directory/
├── agent_binary
└── config.yaml

The config.yaml file specifies command-line arguments per named configuration:

eval: default
configs:
  - default: "--mode standard"  # config used for single config evaluation. 
  - aggressive: "--mode aggressive"

If test_all_configs = true, all listed configurations will be tested. Otherwise, only the one under eval is used.

Repository Structure

.
├── Cargo.toml                 # Crate manifest
├── src/
│   ├── agent_collector/       # Agent compilation and loading logic
│   ├── constraints.rs         # Resource limits enforcement
│   ├── configuration.rs       # Evaluation configuration
│   ├── server.rs              # Core evaluation logic
│   ├── tournament_strategy.rs # Tournament scheduling and formats
│   └── ...                    # Other internal modules
├── README.md
└── TODO.md

Usage Example

use anyhow;
use ai_tournament::prelude::*;
use std::{collections::HashMap, time::Duration};

// Your custom game implementing the Game + GameFactory traits
use crate::YourGame;

fn main() -> anyhow::Result<()> {
    // Define per-agent constraints
    let constraints = ConstraintsBuilder::new()
        .with_ram_per_agent(1000) // in MB
        .with_action_timeout(Duration::from_millis(100))
        .build()?;

    // Define evaluator behaviour
    let config = Configuration::new().with_allow_uncontained(true);

    let factory = YourGame::new(); // Your game logic
    let evaluator = Evaluator::new(factory, config, constraints);

    let tournament = SinglePlayerTournament::new(10); // Run 10 games per agent
    let results: HashMap<String, SinglePlayerScore> = evaluator.evaluate("path_to_agents_directory", tournament)?;

    // Sort and display scores
    let mut sorted = results.iter().collect::<Vec<_>>();
    sorted.sort_by(|a, b| b.1.cmp(a.1));
    for (agent_name, score) in sorted {
        println!("{agent_name}: {score:?}");
    }

    Ok(())
}

[!NOTE] Agents must be Rust crates or precompiled binaries located in the specified directory. Each agent of each match runs as a separate, isolated process.

Example Agent

Here’s a minimal example of an agent compatible with the evaluator system. The agent connects to the evaluator’s server via TCP, reads the game state, and responds with an action:

use std::{
    env,
    io::{Read, Write},
    net::{Ipv4Addr, SocketAddrV4, TcpStream},
    str::{FromStr},
};

use anyhow;

fn main() -> anyhow::Result<()> {
    let mut args = env::args();
    let _ = args.next(); // Skip binary name

    // Read the port number to connect to
    let port = args.next().unwrap().parse()?;
    let addr = SocketAddrV4::new(Ipv4Addr::from_str("127.0.0.1")?, port);
    let mut stream = TcpStream::connect(addr)?;

    // Optionally, reading time_budget and action_timeout from next args
    let total_time_budget = Duration::from_micros(args.next().unwrap().parse()?);
    let action_timeout = Duration::from_micros(args.next().unwrap().parse()?);
    // After the four first arguments (binary name, port number, time budget, and action
    // timeout) will follow your arguments defined in your config file

    let mut agent = YourAgent::new();

    // Interaction loop
    loop {
        let mut buf = [0; 4096];
        let n = stream.read(&mut buf)?;
        let string = str::from_utf8(&buf[..n])?;

        // Parse game state, compute action, send it back
        let game_state = string.parse::<<YourGame as Game>::State>()?;
        let action = agent.select_action(game_state);
        stream.write_all(action.to_string().as_bytes())?;
    }
}

Requirements

• YourGame::State and YourGame::Action must implement FromStr and ToString
• The agent must connect to the provided TCP port and handle communication over the stream
• The agent's select_action call must complete before the action timeout, or it will be forcefully terminated.

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this crate by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.