xcs-rc 1.3.3

Accuracy-based Learning Classifier Systems with Rule Combining

XCS-RC

Accuracy-based Learning Classifier Systems with Rule Combining mechanism, shortly XCS-RC for Python3, loosely based on Martin Butz's XCS Java code (2001). Read my PhD thesis here for the complete algorithmic description.

Rule Combining is novel function that employs inductive reasoning, replacing all Darwinian genetic operation like mutation and crossover. It can handle binaries and real, reaching better correctness rate and population size quicker than several XCS instances. My earlier papers comparing them can be obtained at here and here.

---

Relevant links

• PyPI
• Github repo
• Examples:
  • Classic problems: multiplexer, Markov env
  • Churn dataset
  • Flappy Bird
  • OpenAI Gym

---

Installation

pip install xcs-rc

Initialization

import xcs_rc
agent = xcs_rc.Agent()

Classic Reinforcement Learning cycle

# input: binary string, e.g., "100110" or decimal array
state = str(randint(0, 1))

# pick methods: 0 = explore, 1 = exploit, 2 = explore_it
action = agent.next_action(state, pick_method=1)

# determine reward and apply it, e.g.,
reward = agent.maxreward if action == int(state[0]) else 0.0
agent.apply_reward(reward)

Partially Observable Markov Decision Process (POMDP) environment

# create env and agent
env = xcs_rc.MarkovEnv('maze4')  # maze4 is built-in
env.add_agents(num=1, tcomb=100, xmax=50)
agent = env.agents[0]

for episode in range(8000):
    steps = env.one_episode(pick_method=2)  # returns the number of taken steps

Print population, save it to CSV file, or use append mode

agent.pop.print(title="Population")
agent.save('xcs_population.csv', title="Final XCS Population")
agent.save('xcs_pop_every_100_cycles.csv', title="Cycle: ###", save_mode='a')

Finally, inserting rules to population

# automatically load the last set (important for append mode)
agent.load("xcs_population.csv", empty_first=True)
agent.pop.add(my_list_of_rules)  # from a list of classifiers

---

Main Parameters

XCS-RC Parameters

• tcomb: combining period, number of learning cycles before the next rule combining
• predtol: prediction tolerance, maximum difference between two classifiers to be combined
• prederrtol: prediction error tolerance, threshold for deletion of inappropriately combined rules

How to Set

agent.tcomb = 50 # perform rule combining every 50 cycles
agent.predtol = 20.0 # combines rules whose prediction value differences <= 20.0
agent.prederrtol = 10.0 # remove if error > 10.0, after previously below it

Latest updates

• all related to mutation and crossover is removed
• dependencies like pandas and numpy are removed, as well as data science features

---

Results

Classical Problems: multiplexer and Markov environment:

Flappy Bird from PyGame Learning Environment:

Youtube: CartPole-v0 Benchmark from OpenAI Gym: