tig-innovator

name

tig-innovator

description

Build, test, and optimize algorithms for The Innovation Game (TIG) - a decentralized platform for algorithmic innovation. Use this skill when: (1) Writing CUDA/Rust code for TIG challenges from mathematical derivations (2) Building Docker containers to test TIG algorithms (3) Testing algorithms within the TIG framework (4) Modifying algorithms and benchmarking performance (5) Optimizing hyperparameters across multiple runs (6) Working with TIG challenges: Boolean Satisfiability, Vehicle Routing, Knapsack, Vector Search, Hypergraph, Neural Network Optimizer

TIG Innovator

Build and optimize algorithms for The Innovation Game.

Quick Reference

TIG Docs: https://docs.tig.foundation/innovating/code-submission

Algorithm Structure (in tig-algorithms/src/<challenge>/<algorithm_name>/):

• mod.rs - Rust implementation with solve_challenge() and help()
• kernels.cu - CUDA kernels (GPU challenges)
• README.md - Submission metadata

Docker Testing:

# CPU challenges
docker run -it -v $(pwd):/app ghcr.io/tig-foundation/tig-monorepo/<CHALLENGE>/dev:latest
build_algorithm <name>
test_algorithm <name> <TRACK> <HYPERPARAMETERS>

# GPU challenges (requires NVIDIA GPU + CUDA 12.6.3+)
docker run -it --gpus all -v $(pwd):/app ghcr.io/tig-foundation/tig-monorepo/<CHALLENGE>/dev:latest

Workflow

1. Implement Algorithm from Mathematical Derivation

1. Read the mathematical derivation (PDF or description)
2. Identify the core optimization/update equations
3. Map equations to the optimizer API:
   • optimizer_init_state() - Initialize buffers, set hyperparameters
   • optimizer_query_at_params() - Optional parameter proposal (lookahead schemes)
   • optimizer_step() - Compute weight updates from gradients

CUDA Kernel Pattern (see references/tig-structure.md):

extern "C" __global__ void kernel_name(
    const float* input,
    const int n,
    // hyperparameters...
    float* output
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < n) {
        // Compute output[idx]
    }
}

Rust Launch Pattern:

let kernel = module.load_function("kernel_name")?;
let cfg = LaunchConfig {
    grid_dim: ((n + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK, 1, 1),
    block_dim: (THREADS_PER_BLOCK, 1, 1),
    shared_mem_bytes: 0,
};
unsafe {
    stream.launch_builder(&kernel)
        .arg(&input).arg(&(n as i32))
        // hyperparameters...
        .arg(&mut output)
        .launch(cfg)?;
}

2. Register Algorithm

Update tig-algorithms/src/<challenge>/mod.rs:

pub mod <algorithm_name>;
pub use <algorithm_name> as c00X_aXXX;

3. Build and Test

See references/testing.md for complete Docker workflow.

# Enter container
docker run -it --gpus all -v $(pwd):/app ghcr.io/tig-foundation/tig-monorepo/<CHALLENGE>/dev:latest

# Build
build_algorithm <algorithm_name>

# Test (example for neuralnet_optimizer)
test_algorithm <algorithm_name> default '{"num_hidden_layers": 2, "accuracy_factor": 500}'

4. Benchmark and Optimize

Use scripts/benchmark.py for systematic hyperparameter search:

python scripts/benchmark.py \
    --algorithm <name> \
    --challenge <challenge> \
    --param learning_rate 0.0001 0.001 0.01 \
    --param beta1 0.9 0.95 0.99 \
    --runs 5

5. Submit

Create README.md with required metadata:

# TIG Code Submission

## Submission Details
* **Challenge Name:** <challenge>
* **Algorithm Name:** <algorithm_name>
* **Copyright:** <year> <owner>
* **Identity of Submitter:** <identity>
* **Identity of Creator of Algorithmic Method:** <identity>
* **Unique Algorithm Identifier (UAI):** null

References

• tig-structure.md - Complete code templates and API reference
• challenges.md - Details on each TIG challenge type
• testing.md - Docker setup and testing procedures