ESA is a proposal for automatic problem solving when involving kinects or logic. The proposal is to add to the top layer of the subsumption architecture proposed by R. Brooks a layer that is capable of simple reasoning to select the appropriate strategy. The interaction among layers defines a plan or subplan for solving a problem. The proof of concept uses OpenAI Gym Lunar Lander challenge for demonstration and test of concept. In this code, the abstraction of the physics is captured by a multi layer perceptron, but could use more sophisticated ways of deducting the results of the interactions although for this simulation it can achieve scores as high as 0.92 according to the tests realized. Another MLP is used to predict the score of the next action, but it achieves pretty lower results because this simple MLP could not abstract the penalties by the OpenAI challenge. The score for this MLP achieved 0.53. This is not a problem for the proposal of ESA since it does not expect for all the subsystems to be fully realiable. The result of the interaction of lower subsumption layers and the heuristic-reasoning in the top layer could develop plans for landing the lander, but it struggles when facing the dimensionality problem. For example, a good heuristic for landing is to achieve the target point with very low speed, but it demands several adjustments to the rotation of the vector. The ship may decide that the best strategy for not landing too fast is achieved by not landing at all or using very long horizontal lines to safely reach the ground. The latter is observed in the results of the paper accompaning this work. When using the Lunar Lander challenge as is, the ESA-Auto tries with higher frequency the horizontal approach. Yet, this work presents a interesting way of defining targets and priorities and strategies for a machine to create a plan. It poses the challenge of balancing the weights for each element. This could be explored by the usage of another machine learning system, such as another neural network, to define the weights. The article also proposes other ways of exploring the ideas by extending the concepts of the ESA. Comparing to other available concepts the ESA is a cheap since it uses simple predictions for the next steps and apply a competition of heuristics to select create a plan.
-
Notifications
You must be signed in to change notification settings - Fork 1
The Enhanced Subsumption Architecture is a proposal for solving problems through competing heuristics within the context of Multi-Agent Systems
License
fsan/ESA_OpenAiGym
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
About
The Enhanced Subsumption Architecture is a proposal for solving problems through competing heuristics within the context of Multi-Agent Systems
Resources
License
Stars
Watchers
Forks
Releases
No releases published
Packages 0
No packages published