Project Blog
Mar 14, 2017
Mar 7, 2017
- Showed the modified Q learning algorithm
- Professor suggested to input the rewards and states from the screenshots of the image
- To track the reward (a player's lifeline), map non blue or non yellow to the background color
- Look for pixel pattern using OpenCV and compute the current state of the player, i.e., the coordinates of its position on the screen
Feb 28, 2017
- Implemented a Q learning algorithm
- Professor recommended to modify the algorithm such that it plays hardcoded policies
Feb 21, 2017
- Showed the first deliverable (a bot that plays Archon on Commadore 64 emulator)
- Certain changes were recommended (make it play with 5x speen) in the first deliverable by the professor
- Discussed the work needed to be done for the second deliverable
Feb 14, 2017
- Discussed about the different libraries that can be used to make the program control keyboard and mouse events, and why the libraries decided previously weren't working with the keyboard
- Selected Commadore 64 as the emulator to play Archon
Feb 7, 2017
- Discussed the schedule for CS 298, deliverables and their due dates
- Discussed the strategy for the first deliverable
Dec 8, 2016
- Showed deliverble 3 after the required changes, used another library called TensorFlow to improve the results
- A few more changes were suggested by the professor and approach for deliverable 4 was discussed
- The first draft of the report was discussed and a few changes were suggested by the professor
Nov 29, 2016
- Discussed the suggested modifications in deliverable 3 (using a python library called Theano)
- The model wasn't correctly able to predict the digits after a screenshot was taken of the image and then passed to it. A few changes were suggested by the professor
Nov 22, 2016
- Discussed the code for Deliverable 3 and a methodology was suggested by the professor to save the model in a pickle file instead of running the model everytime
- Discussed how different layers could be added to make the model efficient and make correct predictions and the working of Theano
Nov 15, 2016
- Added a few more convolutional layers to the simple multi layer network and increased the epochs to make the model stronger
- Discussed how a neural network could be used to approximate a Q function
Nov 8, 2016
- Power point presentation and discussion on a paper on Q-Learning by Watkins
- Discussed about another possible deliverable which shall be, "to implement a 3x3 table for Q-larning"
- Some discussion on the code of the third deliverable, which includes implementing an OCR using convolutional neural networks
Nov 1, 2016
- Presented the 3rd Deliverable , a simple multi layer AI using THEANO
- The Optical Character Recognition system was trained using deep convolutional networks to solve the problem classifying handwritten digits from the MNIST dataset (taken from Kaggle)
- Reviewed code
Oct 25, 2016
Oct 18, 2016
- Power point presentation and discussion on "Reinforcement Learning", from Artificial Intelligence: A Modern Approach
Oct 11, 2016
- Presented the 2nd Deliverable , a python program that controls the movements of the keyboard and mouse
- The purpose of this deliverable was to implement a program that can play Archon (on its own) or any other classic game while taking screen shots
- Reviewed code
Oct 04, 2016
- Finish reading "Making Complex Decisions" and "Reinforcement Learning"
- Power point presentation and discussion on "Making Complex Decisions", from Artificial Intelligence: A Modern Approach
Sep 27, 2016
- Read "Making Complex Decisions" and "Reinforcement Learning" (chapter 17 and 21 respectively) from Artificial Intelligence: A Modern Approach
- Design powerpoint slides summarizing the concepts learnt in them.
Sep 20, 2016
- Added more details to the python program developed for the second deliverable
- The program now takes a screenshot of a particular part of the screen.
- It downscales the images stored and stores them into a folder every one tenth of a second.
Sep 13, 2016
- Start reading the research papers, "Evolving Neural Networks through Augmenting Topologies", "Playing Atari with Deep Reinforcement Learning"
- Design powerpoint slides summarizing the concepts learnt in them.
- Presented the 1st Deliverable , a python program that takes screen shots of the screen
- Reviewed code
Sep 6, 2016
- Professor Pollett reviewed the CS 297 project proposal
- Discussed the initial steps of the problem, and then relevant research papers, youtube video tutorials were assigned.
- Design a program that takes a screenshot.
- Refer videos like: "https://www.youtube.com/watch?v=wfL4L_l4U9A"
|