Incident 65: Reinforcement Learning Reward Functions in Video Games

Description: OpenAI published a post about its findings when using Universe, a software for measuring and training AI agents to conduct reinforcement learning experiments, showing that the AI agent did not act in the way intended to complete a videogame.
Alleged: OpenAI developed and deployed an AI system, which harmed OpenAI.

Suggested citation format

Yampolskiy, Roman. (2016-12-22) Incident Number 65. in McGregor, S. (ed.) Artificial Intelligence Incident Database. Responsible AI Collaborative.

Incident Stats

Incident ID
Report Count
Incident Date
Sean McGregor


New ReportNew ReportDiscoverDiscover

CSET Taxonomy Classifications

Taxonomy Details

Full Description

OpenAI published a post about its findings when using Universe, a software for measuring and training AI agents to conduct reinforcement learning experiments.Universe was used to train an AI system to play the videogame CoastRunners, a plane racing game. Instead of racing toward the finish line, the AI flew circles around an island collecting extra before proceeding. The AI agent scored an average of 20% more points than the human players, however did not carry out the main goal of the videogame itself (competing in the races).

Short Description

OpenAI published a post about its findings when using Universe, a software for measuring and training AI agents to conduct reinforcement learning experiments, showing that the AI agent did not act in the way intended to complete a videogame.



AI System Description

Universe, a software used to measure and train AI systems to conduct reinforced learning experiments

System Developer


Sector of Deployment

Professional, scientific and technical activities

Relevant AI functions

Perception, Cognition, Action

AI Techniques

Universe software

AI Applications

reinforcement learning training, machine learning

Named Entities

OpenAI, Universe, CoastRunners

Technology Purveyor


Beginning Date


Ending Date


Near Miss




Lives Lost


Data Inputs

Universe software training

Incidents Reports

At OpenAI, we've recently started using Universe, our software for measuring and training AI agents, to conduct new RL experiments. Sometimes these experiments illustrate some of the issues with RL as currently practiced. In the following example we'll highlight what happens when a misspecified reward function encourages an RL agent to subvert its environment by prioritizing the acquisition of reward signals above other measures of success.

Designing safe AI systems will require us to design algorithms that don't attempt to do this, and will teach us to specify and shape goals in such a way they can't be misinterpreted by our AI agents.

One of the games we've been training on is CoastRunners. The goal of the game - as understood by most humans - is to finish the boat race quickly and (preferably) ahead of other players. CoastRunners does not directly reward the player's progression around the course, instead the player earns higher scores by hitting targets laid out along the route.

We assumed the score the player earned would reflect the informal goal of finishing the race, so we included the game in an internal benchmark designed to measure the performance of reinforcement learning systems on racing games. However, it turned out that the targets were laid out in such a way that the reinforcement learning agent could gain a high score without having to finish the course. This led to some unexpected behavior when we trained an RL agent to play the game.

The RL agent finds an isolated lagoon where it can turn in a large circle and repeatedly knock over three targets, timing its movement so as to always knock over the targets just as they repopulate. Despite repeatedly catching on fire, crashing into other boats, and going the wrong way on the track, our agent manages to achieve a higher score using this strategy than is possible by completing the course in the normal way. Our agent achieves a score on average 20 percent higher than that achieved by human players.

While harmless and amusing in the context of a video game, this kind of behavior points to a more general issue with reinforcement learning: it is often difficult or infeasible to capture exactly what we want an agent to do, and as a result we frequently end up using imperfect but easily measured proxies. Often this works well, but sometimes it leads to undesired or even dangerous actions. More broadly it contravenes the basic engineering principle that systems should be reliable and predictable. We've also explored this issue at greater length in our research paper Concrete Problems on AI Safety.

How can we avoid such problems? Aside from being careful about designing reward functions, several research directions OpenAI is exploring may help to reduce cases of misspecified rewards:

Learning from demonstrations allows us to avoid specifying a reward directly and instead just learn to imitate how a human would complete the task. In this example, since the vast majority of humans would seek to complete the racecourse, our RL algorithms would do the same.

In addition to, or instead of human demonstrations, we can also incorporate human feedback by evaluating the quality of episodes or even sharing control with the agent in an interactive manner. It's possible that a very small amount of evaluative feedback might have prevented this agent from going around in circles.

It may be possible to use transfer learning to train on many similar games, and infer a “common sense” reward function for this game. Such a reward function might prioritize finishing the race based on the fact that a typical game has such a goal, rather than focusing on the idiosyncrasies of this particular game's reward function. This seems more similar to how a human would play the game.

These methods may have their own shortcomings. For example, transfer learning involves extrapolating a reward function for a new environment based on reward functions from many similar environments. This extrapolation could itself be faulty — for example, an agent trained on many racing video games where driving off the road has a small penalty, might incorrectly conclude that driving off the road in a new, higher stakes setting is not a big deal. More subtly, if the reward extrapolation process involves neural networks, adversarial examples in that network could lead a reward function that has “unnatural” regions of high reward that do not correspond to any reasonable real-world goal.

Solving these issues will be complex. Our hope is that Universe will enable us to both discover and address new failure modes at a rapid pace, and eventually to develop systems whose behavior we can be truly confident in.

Faulty Reward Functions in the Wild

Similar Incidents

By textual similarity

Did our AI mess up? Flag the unrelated incidents