Artificial Intelligence Safety And Security Independent Study Week 11

This week I will be doing a comprensive, critical review of open-ended curiousity driven exploration, and other open-ended learning techniques. That's a bit of a vauge subject: AlphaGo could be considered open-ended since it learns to exceed human abilities through self-play. To reduce the search space, I'll resetrict this to situations where the goal is unclear or non-existent, like the real world.

I still haven't done the experiment to run the DL Curiousity algorithm on Game of Life. I need to make time for that, it could be very interesting and certainly a good programming task.

Let's review some papers!

Open-Ended Learning: A Conceptual Framework Based on Representational Redescription

They establish a very rought framekwork of how open-ended learning could take place. The paper seems like a bunch of mumbo jumbo to me, but I also like to see experiments or at least some incomprehensible math before I start believing what a paper says. One valuable insight is the problem formulation: Imagine you are a human, and you are going to deploy a robot into some sort of environment to complete tasks. What kind of tasks? You have no idea. Maybe it has to complete mazes, maybe it has to kill other robots, or play chess, or invent interesting chess variants to keep its fellow robots entertained. The point is, this problem sounds impossible. How can you make a robot that just "does stuff"? However, curiousity seems to be able to perform reasonably well at this task. Not great, but it certainly tries. I don't understand how their "framework" solves this problems, but it concerns greatly with representing the world, which was a large part of curiousity.

Curiosity-driven Exploration by Self-supervised Prediction

This is the one I reviewed before. It seems to be the most powerful exploration algorithm, but it's only been used on a few tasks. I want to do some more tests with it as part of my paper this semester.

I'm going to start looking through their references, to see what previous algorithms were attempted for this task. I'll refer to this agent as Parthak, after the first author.

For more details and termimology, read this blog post

A possibility for implementing curiosity and boredom in model-building neural controllers (1991)

This paper's curiosity reward signal is based on the distance between the network's predicted future state and the actual future state. This is slightly different than Parthak, which uses the FDM to create agent-relavent LSRs. Parthak's IDM is exactly the same as this paper's IDM. I think Parthak's improvements over this are more about computer resources improving since 1991 and less about improvements in technology.


This is a human study, where participants clicked on a blurred image to reveal parts of it. They measured curiousity for this task, the experiment would be an interesting game for an AI to play.

Exploration in Model-based Reinforcement Learning by Empirically Estimating Learning Progress (NIPS 2012)

I don't understand this paper at all. They talk about how their approach can recover from "incorrect priors" and changes in dynamics.

Formal Theory of Creativity, Fun, and Intrinsic Motivation (IEEE 1990–2010)

THIS is really the key paper for creativity. It has a big list of several curiosity AI papers, and provides a one paragraph summary of how human curiosity could work:

For a long time I have been arguing, using various wordings, that all this behavior is driven by a very simple algorithmic mechanism that uses reinforcement learning (RL) to maximize the fun or internal joy for the discovery or creation of novel patterns. Both concepts are essential: pattern, and novelty. A data sequence exhibits a pattern or regularity if it is compressible [45], that is, if there is a relatively short program program that encodes it, for example, by predicting some of its components from others (irregular noise is unpredictable and boring). Relative to some subjective observer, a pattern is temporarily novel or interesting or surprising if the observer initially did not know the regularity but is able to learn it. The observer’s learning progress can be precisely measured and translated into intrinsic reward for a separate RL controller selecting the actions causing the data. Hence the controller is continually motivated to create more surprising data

They also list several existing curiosity algorithms. They tend to work by maximizing the prediction error of a network trying to predict the environment. Parthak isn't much different, except for the use of an environment compressed representation (the latent space representation) that correlates to the agent. It is feasable that you could use the same mechanism to make a reward-relavent LSR.

Unifying Count-Based Exploration and Intrinsic Motivation (NIPS 2016)

Count based learning is used in finite states, so that a state visited often is "boring" and one yet unseen is "interesting". They find a way to create peudo-counts, and show that this is comparable to intrinsic motivation.

Variational Information Maximisation for Intrinsically Motivated Reinforcement Learning (NIPS 2015)

This is an algorithm for mutual information for high-dimensional data. For a small environment, the Blahut-Arimoto algorithm is used but it has exponential complexity. This paper is useful because many AI algorithms in the past have used mutual information, and couldn't scale to pixel space problems.


Innovation Engine

Encouraging Creativity and Curiosity in Robots

Evolution with a curiosity reward

Curiosity-Driven Optimization

Using curiosity to explore a cost surface