Simpler design, while preserving effectiveness.
With this version, Adaptive Neuro-Symbolic Network Agent (ANSNA), derived from Non-Axiomatic Reasoning System theory, has been fully implemented, and a more minimal system, Minimal Sensorimotor Component (MSC), has been derived from it. Feel free to try it on any operating system, or on your phone using Termux. What are its capabilities? Learning from event streams, and reaching goals by subgoal derivations and operation invocations. Differently than Reinforcement Learning, it allows for simultaneous pursuing of multiple goals, and allows goals to change in any moment. (remember: changing the reward function for usual RL agents would demand a complete re-training!) Also it does not assume a markovian environment where the next state depends on a current state and only on the current state (not the past), there are no states in classical sense. But there are events that serve as contextual cues, which, dependent on context (meaning which other events happened, and when), will likely trigger others to appear in a certain time. Invariants, are essentially to be found by the agent through learning and have varying degree of "stability", there are no ideal "states".
Metrics:
<<ANSNA Follow test successful goods=572 bads=72
pong: Hits=18122 misses=6100 ratio=2.970820 time=1223938
pong2: Hits=21580 misses=4298 ratio=5.020940 time=1148410
numeric-pong: Hits=4469 misses=2881 ratio=1.551198 time=379015
numeric-pong2: Hits=2477 misses=4964 ratio=0.498993 time=327395
All tests ran successfully
Repositories:
https://github.com/patham9/ANSNA
https://github.com/patham9/MSC/
Publications:
https://link.springer.com/chapt…/10.1007/978-3-030-27005-6_8
https://link.springer.com/chapt…/10.1007/978-3-319-97676-1_8
https://www.springer.com/gp/book/9781402050442
https://www.worldscientific.com/worldscibooks/10.1142/8665