Artificial Intelligence Can Play Games Like Humans

2020-01-31

Can artificial intelligence play text-based games like humans? This is a question recently studied by the Uber AI Research team. They used a system based on exploration and imitation learning to solve the game using strategies that follow high-reward trajectories. Text-based computer games describe their world to players using natural language and expect players to interact with the game using text. These games are interesting because they serve as a testing ground for AI understanding, problem-solving, and language generation. Additionally, these skills can interact with the environment in a learning setting, rather than being limited to a fixed set of actions like existing methods, which can only solve simple games. One challenge in developing text game AI is how to compete with a large action space. For example, if there are 20,000 words in the vocabulary and sentences can have a maximum of 7 words, the total number of actions would be an astonishing 1.28e^30. The new model maps observed outcomes to actions while keeping track of unexplored areas in the game space. In the first phase, the model explores the environment and records visited locations in archive cells. These cells contain a series of observation data that are mapped to the same representation through mathematical functions. Each cell is associated with metadata, including the trajectory pointing to the cell, the length of the trajectory, and the cumulative reward of the trajectory. In each game session, the model selects a cell based on the metadata and starts random exploration from the end of the trajectory associated with that cell. The goal of the second phase is to transform "fragile" sequences of actions into strategies that can be applied across different game applications, even generalizing to invisible games. In a series of experiments, the team designed three different scenarios: single-player mode, where one model is trained and tested on a game; cooperative mode, where multiple models are trained and tested on a game; and for simplicity, in all games, the maximum number of steps is set to 50. The final results are promising, as the team reports that the new model found an optimal strategy that reduces actions by approximately half compared to previous state-of-the-art systems, while still maintaining a trajectory. The researchers note that although it is not a perfect method, there is significant overlap in the game descriptions, leading to similar observations influencing the strategy. However, compared to previous methods, there has been a breakthrough in finding good development trajectories in large action games.

RECENT AI NEWS

RECENT AI TOOLS