During my A levels I completed an award winning A* EPQ. This is a dissertation equivalent to half an A level. For the topic of my EPQ, I chose to continue a personal project, developing machine learning.
I developed a multi-technique learning AI agent for supervised learning that works on reinforcement tasks.
This works by watching humans play each other, or from creating its own data-set using brute force, or by using an existing data-set.
Once a data-set is created it then uses stochastic gradient descent to train a neural network. When the error of the data-set is low enough, evolution then takes over the learning of the network to allow the agent to improve beyond the starting data-set. This learned all the tasks I tested it on faster with more accuracy and reliability than the other AI agents at the time.
This is one of many reinforcement learning environments I used from OpenAI's Gym
Evolution | Brute Force | Combined | Random | |
---|---|---|---|---|
CartPole | 20 | N/A | 5 | N/A |
Simple Dataset | 3.5 | 6.5 | 2.5 | N/A |
Complex Dataset | N/A | 46 | 28 | N/A |
Tic Tac Toe | N/A | 240 | 150 | N/A |