Summary of the Project: (30 points) Our work focuses on the natural language interface. Most items focus on retrieving user’s commands, analyze iit via AI algorithms, and reflect via agent’s actions in the game. For the framework, it may be provided by professor, or we’ll built the simple framework just enough for demonstrating the AI algorithms via agent. For the Minecraft Project we will focus on natural language interface. The main idea of this project is for our agent to execute actions based on commands entered by the user. The input will consist of the text command and the data structure representing the environment of the world in minecraft. Since we are mostly focusing in classifying the text, we will work with a simple Minecraft world. However, as the project progresses then we will add more complexity to the world. One of the main goals is for the agent to be able to distinguish between a verb and a subject when given a command. It is important for our agent to classify these two components within a sentence in order for it to execute the action. Most of the actions will include the agent to move to different places in the environment and to be able to identify the objects in its surroundings. This will require an algorithm to find the shortest path to the objects. The output will consist of a set of actions that our agent will perform for each command given by the user.
AI/ML Algorithms: (10 points) Build a dictionary of terms that are used in natural language and that are frequently used in commands and orders. For instance, words such as “Go”, ”Move”, ”Forward”, ”Cut” and etc will be included in the dictionary. The agent will use this dictionary to determine if the text given by the user is a command or not. This will be performed by a Naive bayes classifier, where the features will be a set of words from the input(The text given by the user). From these features, the algorithm will predict if the sentence input is a command. After the agent has determined that the sentence is a command, the next step will be to retrieve the verb and the subject from the sentence. For example, if the user inputs, “Cut the tree”, then the algorithm ill retrieve the word “cut” and “tree”. (It could happen that as we progress with this project we can classify other things from sentences such as adjectives. Let’s focus on the “easy” aspects first).Once the verb and the subject have been obtained, the agent will use the second input which is the data structure that represents the Minecraft World and will look for the object index. Finally, using a shortest path algorithm (dijkstra’s algorithm) the agent will move to the object and perform the verb(action). (As I mentioned before, as we progress through the project we can implement other stuff such as adjectives, these adjectives could describe different positions of some objects such as “behind” the house.One way to implement the this could be by finding the shortest distance to the furthest block the object from the agent). The main idea is to sending command to our agent to perform an action. Via AI algorithms applied to the coding, the actions in return will be performed differently from the normal actions, such as paths found have to be the shortest paths, when agent builds something, job has to be done quickly, etc … Due to the limitation in this course, our project is only focus on the AI algorithms in these four mains categories: Find the paths Find items and objects within the environment such as houses and trees Attack quickly/use tool properly
For example: go to the pig: the agent will go to the pig Cut the tree: the agent will find the tree and cut
Classify Actions and objects We first will classify which word is in Action category or Object category. For example: Go to the dog: “Go” would be classified as the Action and “Dog” would be classified as the object. We will first store our object into a dictionary along with it’s index in environment, then we’ll find the way to reach that position fastest using any shortest path algorithm. Shortest path implementation: Tool : we will use natural language tool kit, sklearn, tensorflow.
Evaluation and conclusions:(30 points) Evaluation plan: Quantitative evaluation: This is a set of capabilities that the Agent has to successfully perform.
Qualitative analysis: Performance of the agent will be measured by how well it is able to classify a command. This will be done by creating a dictionary where the key is text and the value is a ‘0’ or ‘1’. 0 = not a command and 1 = command. This Dictionary will be manually built. Also, we will test the performance of our agent by passing a list of all the keys in the dictionary to our agent and our agent will generate its own dictionary with its predictions for each element in the input. Then, we will compare both dictionaries and measure a similarity percentage. Furthermore, the accuracy of classifying a verb and a noun will be tested similarly.