Awesome AI Agents › Generative Agents Simulation
The project 'aleju/mario-ai' is focused on training a deep reinforcement learning model to autonomously play the first level of Super Mario World using raw pixel inputs without any hand-engineered features. It employs deep Q-learning combined with a Spatial Transformer network to predict the best actions based on the game state. The training method uses a replay memory to store sequences of game screens and actions, allowing the model to learn from past experiences by predicting expected action reward values. Inputs to the model include the last four actions represented as two-hot vectors and the last four screenshots downscaled to 32x32 grayscale images, plus the most recent screenshot at 64x64 grayscale. The model outputs expected rewards for pairs of actions, one arrow button and one control button, enabling complex moves like jumping while moving right. The reward function incentivizes moving right, finishing the level, and penalizes death, with a discount factor for future rewards. The error function uses a selective mean squared error to focus training on chosen actions with known rewards. The policy is epsilon-greedy, starting with high exploration and gradually reducing it. The model architecture consists of three branches: action history, screenshot history, and the last screenshot, with convolutional layers and a spatial transformer to focus on areas of interest. The project highlights limitations such as difficulty generalizing to other levels with more complex challenges like jumping puzzles and enemies. It requires a Linux environment with an NVIDIA GPU, CUDA, and CUDNN for training. Overall, this project demonstrates an advanced application of deep reinforcement learning techniques to a classic video game environment, showcasing the integration of convolutional neural networks, spatial transformers, and reinforcement learning policies to achieve autonomous gameplay.
https://github.com/aleju/mario-ai
This project provides a simulation framework for generative agents that mimic believable human behaviors within an interactive game environment, enabling running, saving, replaying, and demonstrating agent-based simulations.
Voyager is an open-ended embodied agent powered by large language models that autonomously explores and learns diverse skills in Minecraft through lifelong learning and an evolving skill library.
SFighterAI is a deep reinforcement learning-based AI agent designed to master and beat the final boss in Street Fighter II: Special Champion Edition using only game screen pixel data.
A comprehensive Godot framework for creating smooth and complex AI motion using steering behaviors in both 2D and 3D games.
WorldX turns a single sentence into a running simulated world, generating maps and characters with LLMs and letting autonomous agents act, remember and form stories inside it.
Control room for LLM agent societies: pause and scrub a running simulation, question individual residents, inject instructions into the next step, and export replayable experiment packs.
PyGame Learning Environment (PLE) is a Python-based reinforcement learning platform that provides a variety of games and tools to facilitate the development and testing of reinforcement learning algorithms.
A locally runnable, low-cost implementation of generative agents simulating human-like behavior in an interactive Dungeons & Dragons setting using large language models.