A weight agnostic neural network performing BipedalWalker-v2 task at various shared weight parameters.

Weight Agnostic Neural Networks

Abstract

Not all neural network architectures are created equal, some perform much better than others for certain tasks. But how important are the weight parameters of a neural network compared to its architecture? In this work, we question to what extent neural network architectures alone, without learning any weight parameters, can encode solutions for a given task. We propose a search method for neural network architectures that can already perform a task without any explicit weight training. To evaluate these networks, we populate the connections with a single shared weight parameter sampled from a uniform random distribution, and measure the expected performance. We demonstrate that our method can find minimal neural network architectures that can perform several reinforcement learning tasks without weight training. On supervised learning domain, we find network architectures that can achieve much higher than chance accuracy on MNIST using random weights.

Supplementary Videos

Below are additional videos that accompany the paper. Please refer to Section 4, Experimental Results of the main text for additional information.

Cartpole Swingup

 

Network and performance chart vs weight at Generation 8.

 
 

Generation 8, weight set to -0.5.

 
 

Generation 8, weight set to +1.5.

 
 

Network and performance chart vs weight at Generation 32.

 
 

Generation 32, weight set to -1.0.

 
 

Generation 32, weight set to +1.5.

 
 

Network and performance chart vs weight at Generation 128.

 
 

Generation 128, weight set to -0.75.

 
 

Generation 128, weight set to +2.0.

 
 

Network and performance chart vs weight at Generation 1024 (champion network).

 
 

Generation 1024, weight set to -2.0.

 
 

Generation 1024, weight set to +1.0.

 
 

Fine-tuned individual weights of champion network (Average score 932 ± 6).

 
 

Bipedal Walker

 
 

Network and performance chart vs weight of Bipedal Walker agent as shown in Figure 1 of the main text.

 
 

The more complicated champion network for BipedalWalker-v2 found in later generations.

 
 

Weight set to -2.0

 
 

Weight set to -1.5

 
 

Weight set to -1.0

 
 

Failure cases at other weight values.

 
 

At some non-optimal weights (here, weight set to +1.14), it performs non trivial actions like balancing.

 
 

Fine-tuned individual weights of champion network (Average score 332 ± 1).

 
 

WANN discovered if we allow connection between outputs.

 
 

Rollout of policy using above network, weight set to -1.0

 
 

Car Racing from Pixels

Champion network for CarRacing-v0.

 
 

Mean Cumulative Reward vs Weight for champion network.

 
 

Weight set to -1.4

 
 

Weight set to +1.0

 
 

Fine-tuned individual weights of champion network (Average score 876 ± 91).

Open Source Code

The instructions to reproduce the experiments in this work is available here.