09 Feb 2018


Training robots to perceive, act and communicate using multiple modalities still represents a challenging problem, particularly if robots are expected to learn efficiently from small sets of example interactions. We describe a learning approach as a step in this direction, where we teach a humanoid robot how to play a game like humans. Attention is now on training the robot to play the game and to interact in this game like humans.

Multimodal deep reinforcement learning agent perceives multimodal features and exhibits verbal and non-verbal actions while playing. Experimental results using simulations show that the robot can learn to win or draw up to 98% of the games.

What Is Reinforcement?

The term reinforce means to strengthen and is used in psychology to refer to anything stimulus which strengthens or increases the probability of a specific response. We all apply reinforcers every day, most of the time without even realizing we are doing it. You may tell your child “good job” after he or she cleans their room; perhaps you tell your partner how good he or she look when they dress up, or maybe you got a raise at work after doing a great job on a project. All of these things increase the probability that the same response will be repeated. In this case, children learn which behaviour is good or bad based on the response it receives along the way.

When better outcomes for the machines are reinforced and adverse choices are discouraged. This is called reinforcement learning.

Even though humans can learn from their mistakes due to the reinforcement response given to them, however, a machine can’t learn from its flaws because it has no emotion. They only recognise and rely on the input data, assignment or mission given to them.

Reinforcement learning has been most successful in very specific, controlled situations that it has been applied. To create machines and programs that are more effectual, then, there is a need to develop a common sense and handle more complex, less structured challenges. In other words, a need to be able to infer when there is a real problem or mission in a living and changing environment.

Likewise, because the real world is a bit complex and can be easily understood by the machines, generative models is a deep reinforcement method that can be used. This mimics the real-world environment, it will generate environmental behaviour and feedback for simulated training settings.  Thus, this can help to further monitor and predict what the machines can do in the real world.

With deep reinforcement learning, Machines will need to learn how to adjust their actions in real life human environments and not just act on clear rules or processed data, but rather on mood and actions, unlike humans who are most times not logical. This will help in other areas like marketing and customer service because they will be able to adjust to customers’ enquiries and complaints, and to make solid and excellent unbiased decision.

This approach allows machines to work out for itself which behaviours are most useful to master. Humans control the learning framework of the simulated environment and let the machine determine optimal behaviour models on its own.


Article by: Busayo Tomoh


Leave a Reply

Notify of