Bookmark this: tool.io/reinforcement-learning/policy-evaluation/source-code
Reinforcement-Learning Policy Evaluator Source Code
Browse the implementation files for Reinforcement-Learning Policy Evaluator.tool.ts typescript
import { parseFiniteNumberList } from '../../parse/numeric-dataset/tool';
import { parseFiniteNumber } from '../../parse/number/tool';
import { parseInteger } from '../../parse/integer/tool';
export const evaluatePolicy = (
rewardsInput: unknown,
transitionInput: unknown,
discountInput: unknown,
iterationsInput: unknown
): string => {
const rewards = parseFiniteNumberList(rewardsInput);
const discount = parseFiniteNumber(discountInput);
const iterations = parseInteger(iterationsInput);
if (
!rewards ||
discount === null ||
discount < 0 ||
discount >= 1 ||
iterations === null ||
iterations < 1 ||
iterations > 100_000 ||
typeof transitionInput !== 'string'
)
return '';
let transitions: unknown;
try {
transitions = JSON.parse(transitionInput);
} catch {
return '';
}
if (
!Array.isArray(transitions) ||
transitions.length !== rewards.length ||
transitions.some(
(row) =>
!Array.isArray(row) ||
row.length !== rewards.length ||
row.some(
(value) =>
typeof value !== 'number' ||
!Number.isFinite(value) ||
value < 0
) ||
Math.abs(
row.reduce((sum, value) => sum + value, 0) - 1
) > 1e-9
)
)
return '';
const typedTransitions = transitions as number[][];
let values = Array(rewards.length).fill(0) as number[];
for (
let iteration = 0;
iteration < iterations;
iteration += 1
)
values = rewards.map(
(reward, state) =>
reward +
discount *
typedTransitions[state].reduce(
(sum, probability, nextState) =>
sum + probability * values[nextState],
0
)
);
if (values.some((value) => !Number.isFinite(value)))
return '';
return JSON.stringify(
{
discount,
iterations,
values: values.map((value) =>
Number(value.toPrecision(10))
)
},
null,
2
);
};
export const Tool: TypeTool = {
name: 'Reinforcement-Learning Policy Evaluator',
slug: 'reinforcement-learning/policy-evaluation',
intro:
'Evaluate a fixed finite-state policy by iterating its reward vector and policy-induced transition matrix.',
tool: evaluatePolicy,
i: [
{
example: '1, 0',
label: 'State Rewards',
placeholder: 'One reward per state',
render: 'textarea',
type: 'text'
},
{
example: '[[0.5,0.5],[0.2,0.8]]',
label: 'Transition Matrix',
placeholder: 'JSON row-stochastic matrix',
render: 'textarea',
type: 'text'
},
{
example: '0.9',
label: 'Discount Factor',
placeholder: 'Value from 0 to less than 1',
type: 'number'
},
{
example: '100',
label: 'Iterations',
placeholder: 'Evaluation iterations',
type: 'number'
}
],
o: [
{
example: '{ "values": [4.19, 2.57] }',
label: 'State Values',
placeholder: 'Policy values will appear here...',
render: 'textarea',
type: 'text'
}
],
cite: [
{
title: 'Markov decision process - Wikipedia',
url: 'https://en.wikipedia.org/wiki/Markov_decision_process'
}
]
};
export default Tool;
Description
Browse the implementation files for Reinforcement-Learning Policy Evaluator.