Assessing Policy Optimization agents using Algorithmic IQ test

This thesis briefly introduces the history and ideology behind the formalised evaluation of intelligence before focusing on the workings of Algorithmic Intelligence Quotient test. As AIQ is closely related to the Reinforcement Learning framework, the following chapter is dedicated to the principles of this framework and introduces some of the popular agents before focusing more in-depth on the agents I have chosen for implementation: Vanilla Policy Gradient and Proximal Policy Optimization. The practical part of this thesis first introduces the history of the prototypical implementation of the Algorithmic Intelligence Quotient and a description of how its code work. This is closely followed by my update of the code base from Python 2 to Python 3, fixes for the code on the Windows operating system, the introduction of a system for logging a failure of an agent and a few other minor tweaks. Next, the thesis focuses on the complexities in introducing agents based on modern Reinforcement Learning architectures into the AIQ test. As AIQ works with environments inverse to the modern standards of OpenAI Gym and its variations, it was necessary to transform them into a state compatible with the AIQ test. The agents are tested to find good default values of a newly introduced parameter before using various statistical methods to compare the new agents against each other and then against originally implemented agents by using the acquired data. Due to its similarity to other existing benchmarks of implemented agents, the results of the comparison between newly implemented agents support the viability of AIQ as an evaluation tool. The results of the comparison to the initially implemented agents give interesting insight into various agents’ intelligence.