Inspired by Twitch: Map Navigation with Crowdsourced Inputs

The power of crowdsourcing has been widely used since they can be easily accessed and the amount of data gives rise to a more complete and complicated set of models. In this project, it is focused on dealing with the problem that not all sources are reliable,selecting, and combining the inputs strongly affect the quality of the decision. A map navigation problem is formulated to investigate methods to wisely combine instructions rather than naively using majority vote. The results have shown that by training with neural network and logistic regression, those algorithms can have a better performance than majority vote. (Output from standard reinforcement learning is used as ground truth here.)

In this application, we further address the task to organize non-omniscient agents to do maze navigation efficiently. The problem formulation is as follows: learn the policy function that takes instructions from multiple agents and outputs a single action.

The methodology to emulate the online map navigation problem is provided in the following picture:

method_poke

The agents are categorized as:

1.Stateless: Random, Going only 1 direction, Gradient following, Gradient following & obstruction avoiding

2. Stateful: BFS, DFS

And the ground truth is provided by the results from value iterations. The overall performance comparison is shown in the following:

regret

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: