Short Update for January

Paper goes over other efforts for self learning AI playing games — states milestones like Deep Blue and Alpha Go

States closest example of anthing doing this contest is DeepMind playing Atari 2600 games

The Text-Based Adventure AI competion is held during the IEEE Conference on computational Intelligence and Games

“The agent has acces only to the natural language description about his surroundings and effects of his actions.”

Used fantasy books and text form decompiled IF games as a source to train the bot.

Bot deals with events either as a battle, inventory mangament, or navigation.

Agent was ested on a total of 50 text based games, beat last years winner in 12 games and did worse in 11 games.

Section II background very broadly: Interactive Fiction, Text-Based Games(games more focued on gameplay than story), Natural Language Processing(NLP), The Text-Based Adventure AI Competition

Paper gives some examples of MUDs(Multi User Dungeons) that have been used in the past to do simmilir task.

Goes over the competition and an overview of how the BYU-Agent worked

Part 3 goes over their agent

“Our agent is characterized by the following features.:

it uses a huge set of predefined command patterns,
obtained by analyzing various domain-related sources; the
actual commands are obtained by suitable replacements;

it uses language models based on selection of fantasy

it takes advantage of the game-specific behaviors, natural
for adventure games, like fight mode, equipment manage-
ment, movement strategy;

it memorizes and uses some aspects of the current play

it tries to imitate human behavior: after playing several
games and exploring the game universe it repeats the most
promising sequence of commands. We treat the result
reached in this final trial as the agent’s result in this game.”

uses word2vec from TensorFlow

Use of LSTM neural networks

commands were gathered from walkthroughs, tutorials, and decompiled games

“The algorithm uses 5 types of command generators: battle
mode, gathering items, inventory commands, general actions
(interacting with environment), and movement.”

Battle mode: perfers using battle actions. may use actions more than once even if action failed. fights may be based on dice rolls.

Inventory mode: In a new area the AI searches for the room descritopn for items. If the AI takes a new item new commands will be used with the item.

Exploration mode: The AI maps the locaton by using the first few words in a discription. Checks new areas by running movement commands to find new rooms.

Failing commands: If a command seems to be failed by the output text that command is blacklisted at the current locaiton until new items are colected

Restarts: When the AI dies it restarts the game and remembers the steps taken to die. It also know the path taken to increase it’s score.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s