Reinforcement Learning – From Playing Games to Driving Cars
Computers are making use of a seemingly simple approach called reinforcement learning to not just excel at complex games but making driverless cars possible too.
- Jeronimo De Leon
- 14 Feb
What is Reinforcement Learning?
It is an approach to artificial intelligence that gets computers to learn like people by doing things incrementally well, without explicit instruction. The author gets a taste of this at a Barcelona AI conference where self-driving cars in a simulation learn to maneuver through a complicated four-lane highway by practicing moves over and over. This is also the approach that helped AlphaGo, designed by Alphabet subsidiary DeepMind beat the World No. 1 Go player in 2017 – Go is a notoriously complex game, making this achievement particularly remarkable.
Reinforcement Learning – A Historical Perspective
Psychologist Edward Thorndike first documented this principle over 100 years ago with an experiment involving cats who learned to escape from a box after stepping on a lever by chance. An early reproduction in machines was in 1951, where a machine simulating a rat escaping from a virtual maze saw some synaptic connections strengthen to reinforce successful behavior. Following decades saw intermittent successes, but using this for more complex tasks became computationally impractical.
The Breakthrough of Deep Learning
In recent years, reinforcement learning has become so formidable because of breakthrough in deep learning, which uses a large simulated neural network to recognize data patterns. This makes storing values for every move and updating them – the basis of reinforcement learning – easier.
This breakthrough is also what made DeepMind first excel phenomenally in Atari video games in 2013, resulting in its acquisition by Google. Today, Google is using these capabilities to make its data centers more energy efficient.
Reinforcement Learning in Self-Driving Cars
This approach is particularly well-suited for simulating humanlike behavior in self-driving cars. Today’s driverless cars often falter in complex situations that involve interacting with human drivers, such as traffic circles. To prevent extreme behavior – taking unnecessary risks or being too hesitant – these cars will have to acquire nuanced skills.
Mobileye, the Israeli firm that developed the Barcelona demo, plans to test the software on a fleet of vehicles with BMW and Intel this year. Google and Uber are already testing reinforcement learning for their vehicles.
While reinforcement learning helps automated driving by enabling good sequences of decisions, which is more efficient than pre-programming all such decisions, there are challenges too. The huge amount of data, the time needed to practice simulations, multiple objectives (avoiding accidents vs. keeping roads safe for other cars) – all make this an extremely complex area.
Discover, Learn and Evaluate AI Companies and Solutions
The latest updates from AI companies in your industry
Get a weekly newsletter with the latest posts directly from the AI companies. Follow companies to tailor your feed.
JEDAI PRODUCT SUITE
Anagog provides a solution suite that offers to mobile app owners the most cost-effective, end-to-end, customer engagement solution.
Experience the leading AI demand forecasting platform.
SPELL - HYPERPARAMETER SEARCHES
Optimize hyperparameters to improve the accuracy of your model with the Spell Hyper command
SPELL - JUPYTER NOTEBOOK SERVER - GPU POWERED
Collaborative Jupyter Notebook or JupyterLab workspace server with powerful GPUs
SPELL - DEEP LEARNING PLATFORM
The fastest and most powerful end-to-end platform for machine learning and deep learning.