https://github.com/seanwevans/damnati
A CUDA-accelerated iterated prisoner's dilemma arena
https://github.com/seanwevans/damnati
arena cuda iterated-prisoners-dilemma prisoners-dilemma tournament
Last synced: about 1 month ago
JSON representation
A CUDA-accelerated iterated prisoner's dilemma arena
- Host: GitHub
- URL: https://github.com/seanwevans/damnati
- Owner: seanwevans
- Created: 2025-09-14T16:28:48.000Z (9 months ago)
- Default Branch: main
- Last Pushed: 2025-09-22T15:15:08.000Z (9 months ago)
- Last Synced: 2025-09-22T17:26:51.247Z (9 months ago)
- Topics: arena, cuda, iterated-prisoners-dilemma, prisoners-dilemma, tournament
- Language: C++
- Homepage:
- Size: 383 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Damnati
Damnati is a CUDA-accelerated tournament simulator for the Iterated Prisoner's Dilemma (IPD). It pits a population of strategies against each other, including classical IPD strategies and optional N-gram learning agents.
## Features
- Runs all pairwise matches between agents on the GPU.
- Supports classic strategies such as **Tit for Tat**, **Grim**, **Pavlov**, and others.
- Optional N-gram learners with adjustable depth, exploration rate, and forgiveness parameters.
- Reports aggregate statistics and per-strategy performance after the tournament.
## Building
The program requires the NVIDIA CUDA toolkit. Build the executable with `nvcc`:
```bash
nvcc -O3 -arch=sm_86 damnati.cu -o damnati
```
## Usage
Run the simulator and adjust parameters using command line flags:
```bash
./damnati --agents 512 --rounds 200 --seed 42 --p-ngram 0.6 --depth 3 --epsilon 0.1 --gtft 0.1
```
### Options
- `--agents N` – number of agents in the population (default: 256)
- `--rounds R` – number of rounds per match (default: 200)
- `--seed S` – RNG seed (default: 1234567)
- `--p-ngram F` – fraction of agents using the N-gram strategy (default: 0.5)
- `--depth D` – N-gram history depth (default: 3, maximum: 15)
- `--epsilon E` – exploration rate for N-gram learners (default: 0.1)
- `--gtft P` – forgiveness probability for Generous Tit for Tat (default: 0.1)
Depths above 15 are rejected to keep GPU memory allocations within practical
limits.
Passing an unknown option will print the usage information and exit with a
non-zero status, so double-check flag names.
## Output
The simulator prints summary statistics in JSON format, including average score, minimum and maximum scores, standard deviation, and per-strategy results. The configuration header now reports the probability used by **Generous Tit for Tat** as `"gtft"`. For example:
```json
{"agents":256,"rounds":200,"p_ngram":0.600,"depth":3,"epsilon":0.100,"gtft":0.100,
"avg_score":1.234, "min":-10, "max":15, "stdev":2.468,
"by_strategy":{...}}
```
The top few agents and their scores are also listed.
## Testing
Unit tests use the [Catch2](https://github.com/catchorg/Catch2) framework and can be
compiled with `nvcc`. From the repository root run:
```bash
nvcc -std=c++17 tests/test_core.cu -o tests/test_core
nvcc -std=c++17 tests/test_cli.cu -o tests/test_cli
```
Execute each binary to run the associated suite:
```bash
./tests/test_core
./tests/test_cli
```
`test_core` exercises the strategy decision logic in `choose_action`, the payoff
accumulation performed by `ngram_update`, and the 64-bit integer square root helper.
`test_cli` verifies that command-line validation reports errors for malformed depth
values and unknown options.